Reasoning model releases

TypeSafe AI's first "System One" model, released Sep 15 2026 in early access. Jev is built for automation workflows rather than chat or copiloting: it maps unstructured state to typed probabilistic decisions and emits parallel structured outputs (JSON / tool calls), trained with what TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD). The company frames it as a "frontier-intelligence function call" — unstructured input in, calibrated decisions out — aimed at agents, classification, and tool use. Proprietary with conditional commercial use; weights are not released and the context window was not disclosed at launch. Served on a single TypeSafe AI serverless route priced at $0.042 per 1M input tokens with output free at launch. Not intended for vision or document-understanding workloads. Figures are vendor-reported and unverified by an independent harness at launch.

Salesforce's first CRM reasoning model, announced Sep 15 2026 (at Dreamforce) and built in collaboration with NVIDIA for the Agentforce platform. Koa was created by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset modeled on nearly three decades of CRM deployments — no customer data was used; the training corpus simulates enterprise reasoning, tool use, and decision-making across 14+ industries (manufacturing, financial services, healthcare, travel, and more). Post-training used Supervised Fine-Tuning plus reinforcement learning (GRPO) with NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel. On Salesforce's own CRM benchmark — real tasks like updating an opportunity, routing a case, or scheduling a follow-up — Salesforce reports Koa matches or exceeds leading models on CRM actions with roughly 3x fewer errors. Salesforce controls the model weights and runs post-training and inference entirely within its own trust boundary (weights not released), offering it as a Salesforce-hosted option for Agentforce. Text/agent reasoning with tool use; no public per-token pricing. Available to select pilot customers at launch, with general availability expected winter 2026 in U.S. regions. Benchmark figures are vendor-reported.

Moonshot AI's mid-tier coding and agentic model, released Sep 11 2026 and positioned between the flagship Kimi K3 and the older Kimi K2.7 Code. Moonshot describes its coding and agent performance as close to K3 with significantly more efficient "thinking," and brings the K3-series thinking controls — adjustable reasoning effort set to low, high, or max, with max as the default. Multimodal input (text, images, and video) with text output, and a 1M-token context window made available across all Kimi membership tiers rather than gated to upper plans as it was when K3 first opened 1M context. Served through Kimi Code under the model id kimi-for-coding, so existing clients pick it up without config changes. Proprietary and preview-status at launch; parameter count and architecture are undisclosed for this specific checkpoint, though it sits within Moonshot's MoE Kimi K2 line. Benchmark and efficiency figures are vendor-reported and unverified by an independent harness at launch.

Shanghai AI Laboratory's first ATRIA-line model, released weights-first on 2026-09-11: the checkpoint and code appeared on GitHub / Hugging Face with no announcement, and a 140-author technical report followed roughly three days later, inverting the usual paper-first convention. Atria Dawn Preview is a 744-billion-parameter agentic Mixture-of-Experts model built on GLM-5.2, designed for long-horizon research agents — carrying scientific work from a method described in the literature through to executable experiments, reproducible metrics, and a report others can inspect. It ships with a 256K-token context under an MIT license with open weights, so it can be self-hosted (documentation and evaluations are published at atria-asi.ai; the checkpoint is on Hugging Face as atria-asi/atria-dawn-preview). On the lab's own table of 16 benchmarks it reports the highest listed score on five tasks, including AutomationBench 53.8, BrowseComp 92.5, DeepSearchQA 96.0, BFCL v4 77.0 and CyberGym 86.5. All figures are vendor / self-reported and were not independently verified at launch.

The higher-performance tier of Sakana AI's Fugu orchestration family, released September 11 2026. Rather than a single monolithic model, Fugu is a learned multi-agent orchestration system — a language model trained to route tasks across a fixed, swappable pool of open-weight and specialized models and to recursively call instances of itself — detailed in Sakana's ICLR 2026 TRINITY (evolved LLM coordinator) and Conductor (RL-discovered natural-language coordination) papers. v2 prioritizes answer quality on complex multi-step reasoning, autonomous research, and full-stack software development, and notably does not include any proprietary frontier model in its pool: no Claude Fable 5, Fable 5.1, or GPT-6 Astra, routing instead over open-weights and specialized models including NVIDIA Nemotron via partnership, which Sakana positions as resilience against single-vendor and export-control risk. Supports configurable reasoning effort (high, xhigh, max), function calling, structured outputs, image and PDF input, and built-in web search. Pricing is $5/$30 per Mtok (cached input $0.50), with a premium tier of $10/$45/$1.00 above 272K-token context. Vendor-reported: 48.3 on Chartography (a visual reasoning / data-interpretation benchmark) versus 27.3 for Anthropic's Opus 5 and 29.5 for Fable 5 — achieved without either closed model in its pool — plus 74.3 on DeepSWE. Parameter count and context ceiling are undisclosed, the figures describe an orchestrated system rather than one set of weights, and all are unverified at launch.

A new cost-optimized tier of Sakana AI's Fugu orchestration family, launched alongside Fugu Ultra v2 on September 11 2026. Like the rest of the family it is a learned multi-agent orchestration engine that routes tasks across a swappable pool of open-weight and specialized models (built on Sakana's ICLR 2026 TRINITY and Conductor methods) rather than a single monolithic model, avoiding dependence on any one proprietary frontier model. Sakana is explicit that Max and Ultra v2 are not separate products but the same core orchestration architecture pointed at different missions — Ultra v2 chases peak capability, Max chases cost. Fugu Max is priced at $2/$6 per Mtok input/output with cached input at $0.25, undercutting the output cost of frontier models such as Claude Sonnet 5, GPT-5.6 Terra, and Kimi K3 by roughly 40-60%. On Sakana's self-reported evaluations it claims the best overall score across six benchmarks including Terminal-Bench 2.1, GPQA-Diamond, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. Parameter count, architecture detail, and context ceiling are undisclosed; figures are system-level and unverified at launch.

The natively multimodal member of Ant Group inclusionAI's Ling-3.0 line, released Sep 10 2026 with MIT open weights (inclusionAI/Ling-3.0-flash-VL). A 124B-total / ~5.5B-active sparse MoE that inherits Ling-3.0-flash's language, reasoning, and long-context ability and extends it with native image and video understanding: a ViT visual encoder feeding a two-layer MLP projector, VideoRoPE positional encoding for video, and the family's 42-layer hybrid backbone alternating Kimi Delta Attention and Gated MLA layers at a 5:1 ratio. Carries a 256K-token context. inclusionAI frames the vision work around three axes — understanding complex visual information, reasoning over visual evidence, and interacting with interfaces (GUI agents) — and reports 42 on the Artificial Analysis Intelligence Index v4.1.1, four points above text-only Ling-3.0-flash at 38. Served free at launch via OpenRouter (inclusionai/ling-3.0-flash-vl) alongside self-hosting on vLLM. Vendor figures, unverified independently at launch.

DeepSeek's efficient, very-low-cost flagship released September 10 2026, retiring V4-Flash and taking over V4-Pro API traffic on September 14 (DeepSeek reports it beats V4-Pro on performance, cost, speed, and total time). A 552B-total-parameter multimodal sparse Mixture-of-Experts model built on a new Causal Encoder-Decoder architecture that activates only ~8B parameters per token on input and ~16B on output for cheaper long-context prefill, with a 1M-token context, up to 384K output tokens, and native image understanding (vision in, text out). The other headline change over July's V4-Flash is memory: FP4 quantization plus "pure CSA2" cross-layer attention reuse compress the KV cache to roughly 890 bytes per token — about an 8x reduction, cutting HBM to a quarter and SSD to an eighth for an equivalent conversation state, which is what makes million-token agentic runs practical on a single node. Open-weight under the MIT license, downloadable and self-hostable, and served across many inference providers. API pricing is $0.30/$1.20 per Mtok input/output at peak (01:00-04:00 and 06:00-10:00 UTC weekdays) and half that off-peak ($0.15/$0.60), with cache hits around $0.006/$0.003 per Mtok. DeepSeek reports it narrowly edges Claude Opus 5 and GPT-5.6 Sol on DeepSWE, but there is no independent benchmark table at launch and vendor performance claims are unverified.

The lightweight multimodal MoE member of Nex AGI's Nex-N2.5 family (Sep 8, 2026). 35B total / 3B active sparse MoE with a 262K context, sharing the Pro model's multimodal agentic stack (computer use, web browsing, visually grounded workflows) at much lower inference cost. Supports vision, reasoning, function calling, and tool use. Apache-2.0 weights on Hugging Face (nex-agi/Nex-N2.5-mini); hosted as a free preview on OpenRouter at launch.

Nex AGI's trillion-scale, text-only MoE agentic model in the Nex-N2.5 family (Sep 8, 2026). 1.6T total / 49B active sparse MoE with a native 1,048,576-token (1M) context. Supports reasoning, function calling, and tool use; no vision or multimodal inputs. Apache-2.0 weights on Hugging Face (nex-agi/Nex-N2.5-Max); no tracked hosted API route at launch.

Nex AGI's flagship multimodal MoE agentic model in the Nex-N2.5 family (Sep 8, 2026). 397B total / 17B active sparse MoE with a 262K context. Builds on Nex-N2 multimodal foundations with stronger computer use, web browsing, and visually grounded agent workflows; supports vision, reasoning, function calling, and tool use. Apache-2.0 weights on Hugging Face (nex-agi/Nex-N2.5-Pro); hosted as a free preview on OpenRouter at launch.

OpenBMB's on-device flagship, released Sep 7 2026 with open weights under Apache 2.0. A 2.52B-parameter (2,516,756,480) dense model on a standard Llama architecture with a 131,072-token context, aimed at strong reasoning and agentic behavior at a size small enough to run locally. OpenBMB reports a 53.9 average across a 34-benchmark comparison set — ahead of the 51.1 it shows for Qwen3.5-4B — with standout math results of 86.5 on both AIME 2025 and AIME 2026, 63.8 on HMMT February 2026, and 94.6 on MATH-500, positioning it at the top of the sub-4B open class. Shipped alongside its training data (including UltraData-SFT-Agent-2609 with 500K agent samples and UltraData-RL-2609 with 80K+ RL samples) and a family of deployment builds — base, mid-training and SFT-only checkpoints, GGUF and MLX conversions, a 4-bit GPTQ version, and a MiniCPM5-2B-DSpark draft model for speculative decoding — which are packaging variants of this release rather than separate models. Vendor-reported figures, unverified independently at launch.

iFLYTEK / SparkLLM's cloud flagship Mixture-of-Experts model in the Spark X2.5 family (announced Sep 7, 2026). 293B total / 30B activated parameters with a 256K (262,144-token) context; multilingual text generation across 200+ languages, focused on coding and agent capabilities with continued math/understanding gains, trained and served on domestic Chinese compute via the xfyun MaaS platform. MaaS list pricing is published in CNY only (input CNY1.6 / 1M, cache-hit CNY0.24 / 1M, output CNY6 / 1M); no first-party USD table, so token prices are left blank here. Distinct from the open-weight Apache-2.0 edge siblings Spark-X2.5-4B and Spark-X2.5-1.7B.

The lightweight member of inclusionAI's (Ant Group) LLaDA2.2 diffusion-language-model family (released Sep 5, 2026). A 16B-total / ~1.4B-active MoE diffusion LLM with a 131,072-token (128K) context and the same Levenshtein-Editing agentic stack as LLaDA2.2-flash at lower inference cost. Apache-2.0 weights on Hugging Face (inclusionAI/LLaDA2.2-mini). Text-diffusion LLM (not an image model).

The higher-quality reasoning tier of OpenAI's GPT-6 Astra, released Sep 4, 2026 (API id gpt-6-astra-pro). It is the same underlying model as GPT-6 Astra served with reasoning.mode set to 'pro', trading latency for higher-quality responses on the hardest professional, coding, research, computer-use, and agentic tasks. Carries a 1.05M-token context window and up to 128K output tokens. Standard pricing is $10/$50 per Mtok input/output (cached input $1/Mtok, batch at half price), with a Fast mode at ~2x the standard rate. Access is limited to ChatGPT Pro, Business, and Enterprise users (and the pro reasoning mode on the API); off by default at launch and enabled per workspace. Shares GPT-6 Astra's positioning around state-of-the-art computer/browser use, finished professional artifacts, long-session Codex coding, and cybersecurity capability that crosses the 'Critical' threshold of OpenAI's Preparedness Framework — so exploit-creation capabilities ship gated behind OpenAI's Daybreak program.

A health- and medicine-domain-tuned variant of Ant Group inclusionAI's Ling-3.0-flash, launched September 4 2026 (model id inclusionai/ling-3.0-flash-sante). Same efficient sparse Mixture-of-Experts base — 124B total parameters, ~5.1B active per token — post-trained for medical knowledge reasoning, clinical safety, evidence-based retrieval, and long-horizon medical tasks, while inclusionAI reports it retains strong general reasoning, coding, and agentic ability. Text-in / text-out only (no vision), with a 262,144-token (256K) context and up to 32,768 output tokens; supports reasoning and tool / function calling. Available first via hosted serverless API — Novita, OpenRouter (inclusionai/ling-3.0-flash-sante), and Vercel AI Gateway — with a time-limited free window at launch (free through Oct 4 on Vercel AI Gateway). Positioned as a developer API for research, retrieval, summarization, and workflow assistance, explicitly not a medical device or a substitute for clinical judgment; no public benchmark table at launch, so treat domain claims as unverified. Like the Fin variant, Sante-specific open weights were not confirmed posted at launch (the base Ling-3.0-flash family ships open-weight, MIT), so treat the open-weight status as announced-but-unconfirmed for this variant.

OpenAI's new frontier flagship, released Sep 3, 2026 (API id gpt-6-astra) as the successor to GPT-5.6 Sol and OpenAI's most capable and most aligned model to date. Positioned around three areas: state-of-the-art computer/browser use, a step change in producing finished professional artifacts (documents, slides, spreadsheets, and websites via Sites in ChatGPT that follow the user's templates), and a jump in cybersecurity capability that crosses the 'Critical' threshold of OpenAI's Preparedness Framework. It decides when to ask clarifying questions versus proceed on assumptions, holds constraints across mid-task steering, and — with an updated Codex harness — keeps searchable notes across context windows instead of lossy compaction. Carries a 1M-token context window; available as gpt-6-astra in the OpenAI API and on Amazon Bedrock (with Microsoft Azure), plus a GPT-6 Astra Pro tier for ChatGPT Pro, Business, and Enterprise (off by default at launch, enabled per workspace). Standard API pricing is $10/$50 per Mtok input/output, with a Fast mode at ~2x price (~$20/$100) for up to 2.5x speed, plus separate cache rates. Vendor-reported benchmarks include OSWorld 2.0 72.6% (at ~47% less time per task than Sol), FrontierMath Tier 4 v2 97.6%, GPQA Diamond 96.0%, Terminal-Bench 4.0 57.7%, and ExploitBench 100%; the marquee ARC-AGI-3 ~99.9% depends on a stateful provider-adapter harness (stateless runs score far lower), and it trails Claude Fable 5.1 on Humanity's Last Exam with tools (57.2% vs 65.0%). Because it crosses the Critical cyber threshold, exploit-creation capabilities ship gated behind OpenAI's Daybreak program; OpenAI also flags a regression in chain-of-thought monitorability as a research priority.

HUMAIN's Arabic-focused frontier MoE model (428B total parameters), commissioned by HUMAIN and delivered on the MiniMax-M3 lineage with HUMAIN Arabic post-training. Natively multimodal (text plus image and video understanding) with agentic tool use and controllable reasoning modes. Available as a limited / research preview on HUMAIN Node (API id humain-m3). Distinct from the live minimax-m3; proprietary weights not yet released.

The flagship sparse-MoE checkpoint of MBZUAI/IFM's K2 Horizon family (released Sep 3, 2026). 375B total / 23B active per token with a native 512K (524,288-token) context and Apache-2.0 weights on Hugging Face (IFM/K2-Horizon-375B-A23B). Ships with k2_horizon reasoning and tool-call parsers for agentic tool use, terminal workflows, and long-horizon reasoning.

Google DeepMind's fast, cost-efficient Flash model released Sep 2, 2026 — its fourth Flash model in under four months and the successor to Gemini 3.7 Flash. At high reasoning it scores 59 on the Artificial Analysis Intelligence Index (up 3 points from Gemini 3.7 Flash and level with sub-maximum efforts of GPT-5.6 Sol and Grok 4.6), 57 at medium (matching GPT-5.6 Terra and Muse Spark 1.2), and 52 at low (matching Gemini 3.6 Flash at ~30% lower cost per task). The improvement is driven mainly by agentic evaluations — t^3-Banking tool use (+12 points to 45%), Terminal-Bench v2.1 coding, and GDPval-AA v2 real-world tasks. It keeps a 1M-token context window and multimodal input (text, image, video, speech) with text output. Pricing matches Gemini 3.7 Flash's current discounted rate of $0.75/$3.75 per Mtok input/output through the end of 2026 ($1.50/$7.50 at standard pricing), with cached input keeping a 90% discount; a ~30% rise in average output tokens per task (to ~48k) lifts cost per task to ~$0.58 at high reasoning despite unchanged per-token pricing. Available in the Gemini app for AI Pro and Ultra subscribers, AI Mode, and Gemini in Google Sheets, and for developers via Google Antigravity, AI Studio, and the Gemini API. Benchmark figures are vendor/third-party-reported.

Meta's successor to Muse Spark 1.2, released Sep 2, 2026. A multimodal reasoning model built for long-running agentic, multi-agent, and coding workflows: it is designed to keep track of information across extended tasks, work through conflicting inputs, and request clarification or confirmation when needed, with an emphasis on concise execution. Accepts text and image input over a 1M-token (1,048,576) context and returns text. Standard-tier API pricing is $1.25 / $4.25 per 1M input/output tokens ($0.15 cached input); as with prior Muse Spark releases a lower-cost muse-spark-1.3-contributor tier is offered in exchange for permission to train future Meta models on prompts and completions. Hosted by Meta and served via OpenRouter (meta/muse-spark-1.3).

An updated snapshot of Alibaba's flagship Qwen3.8-Max, released Sep 2 2026 (model id qwen3.8-max-0902; alias qwen3.8-max-2026-09-02). A post-training upgrade on the unchanged 2.4-trillion-parameter Mixture-of-Experts base (~95B active per token), it sharpens coding for engineering-scale projects and long-horizon autonomous development, multi-tool agent orchestration, and native vision (chart reasoning, document parsing, multimodal perception). Accepts text, image, and video input and returns text, with a 1M-token context, up to ~131K output tokens, and an optional thinking mode carrying up to 256K chain-of-thought tokens. Pricing is unchanged from Qwen3.8-Max at $2.00 / $6.00 per 1M input/output tokens. Vendor-reported gains include a +22-point CodeArena jump to 1,691 (first on that leaderboard at launch); figures are unverified. Served API-first (Alibaba, OpenRouter); the -Max flagship tier is closed-weight, so treat this snapshot as proprietary despite the open-weight base family. Distinct from the base Qwen3.8-Max (2026-08-03).

Anthropic's most capable widely released model, launched Sep 1, 2026 (API id claude-fable-5-1) as a direct upgrade to Claude Fable 5 for agentic coding, long-running problem-solving, and knowledge work. Anthropic reports gains across coding, computer use, multidisciplinary reasoning, and finance/analysis while producing more concise plans and summaries, and cites benchmark results including Terminal-Bench 4.0 55.8%, Terminal-Bench-Science 0.1 52.6%, Humanity's Last Exam 65.0% (with tools), and CursorBench 3.2.0 73.4%. Base pricing is unchanged from Fable 5 at $10/$50 per Mtok input/output, but cache reads were cut 75% to $0.25/Mtok, reducing typical workload cost ~25% (up to ~45% for highly agentic work). Carries a 1M-token context and up to 128K output tokens; available on the Claude API, AWS, Google Cloud, and Microsoft Azure. Cybersecurity safeguards were loosened to allow defensive vulnerability discovery (Claude Code sees ~60% fewer false-positive interventions), while dual-use cyber and life-sciences R&D requests still route to Opus models. First Anthropic release to carry the EU AI Act text watermark.

The latest diffusion large language model (dLLM) from Inception, released Aug 31, 2026. Instead of generating tokens sequentially, Mercury 2.5 produces and refines many tokens in parallel, reaching ~1,107 tokens/sec on standard GPUs. Inception positions it as the fastest reasoning LLM, reporting a 10+ point intelligence gain over Mercury 2 and quality comparable to cost-optimized frontier models (GPT-5.6 Luna Low, Gemini 3.5 Flash-Lite, Claude Haiku 4.5). It supports tunable reasoning levels, parallel tool calls, and schema-aligned JSON output, and targets latency-sensitive production workloads such as search agents, voice pipelines, and coding subagents. 260K-token context, up to 65,536 output tokens; API-only (proprietary) via Inception and OpenRouter (inception/mercury-2.5-preview). List pricing is $0.20/$0.75 per Mtok input/output (a launch promo ran at $0.04/$0.15). Architecture recorded as unknown: it is a diffusion LM rather than a standard dense/MoE autoregressive transformer.

Tencent Hunyuan's fourth-generation flagship, released and open-sourced under Apache-2.0 on Aug 28 2026 (weights and an FP8 variant on Hugging Face as tencent/Hy4-preview). A Mixture-of-Experts model with 770B total parameters and ~49B activated per token: a 78-layer backbone with 256 routed experts plus a shared expert in most layers (8 routed experts selected per token) and a native multi-token-prediction layer for speculative decoding. Carries a 1M-token (1,048,576) context and is aimed at long-horizon software engineering (planning, debugging, verification across extended tasks with tool calling), office and financial analysis, cross-document work, game-development prototyping, and scientific workloads. Available via Tencent Cloud TokenHub and OpenRouter (model id tencent/hy4-preview, tool calling + structured outputs) or self-hosted on vLLM/SGLang; OpenRouter listed launch pricing of $0.834 input / $2.501 output / $0.042 cache-read per 1M tokens (Aug 28 snapshot). In Tencent's internal blind test (163 experts over 203 engineering tasks) Hy4 averaged 2.99 vs 2.92 for GLM-5.3 and 2.94 for Kimi K3 — narrow margins, and Tencent flags preview-stage tendencies to over-reason and over-verify. All figures vendor-reported and unverified by independent labs at launch.

A finance-domain-tuned variant of Ant Group inclusionAI's Ling-3.0-flash, launched August 27 2026. Same efficient sparse Mixture-of-Experts base — 124B total parameters, ~5.1B active per token — post-trained on high-quality financial data (developed with financial institutions and domain experts) for real-world investment and banking workflows: annual reports, financial workbooks, multi-document research, information retrieval, investment analysis, and valuation modeling, with an emphasis on complex multi-step tasks and long-horizon planning. Retains a 262,144-token (256K) context and up to 32,768 output tokens, and supports tool / function calling (tools and tool_choice), though it does not enforce structured JSON output (no response_format). inclusionAI reports it preserves strong general reasoning, coding, and math ability alongside the finance gains, citing finance benchmarks such as FinFIRST, FinSearchComp, and SpreadsheetBench — vendor claims, unverified independently. Launched hosted-API-first with a one-month free window through OpenRouter (inclusionai/ling-3.0-flash-fin); the promised open weights followed as announced and are now on Hugging Face under MIT (inclusionAI/Ling-3.0-flash-Fin, confirmed posted by 2026-09-04, with third-party hosting on DeepInfra and community GGUF quantizations).

The productionized, managed Qwen Cloud API model announced Aug 26 2026 alongside the open-weight Qwen3.8-Flash-Next, with official QwenCloud pricing confirmed Aug 27-28. It runs the same Qwen4-preview architecture as Flash-Next — a 125B-total sparse Mixture-of-Experts that activates ~6B parameters per token (~180B stored once a 51B n-gram embedding table and a multi-token-prediction module are counted; roughly 95% sparsity), built on Gated-DeltaNet + Qwen Sparse Attention with a gated residual stream and Muon-trained large linear layers — but ships as the hosted service rather than the self-host weights. On QwenCloud it defaults to a 1M-token context with built-in tools and accepts text/image/video in, returning text out. List pricing is $0.15 input / $0.47 output / $0.016 cache-hit per Mtok (domestic China Y0.8/Y2.7/Y0.1), roughly a third of DeepSeek-V4-Flash and about one-thirteenth of the Qwen3.8-Max flagship. Also reachable through OpenCode Go's flat-rate subscription. Distinct catalog row from the open-weight Qwen3.8-Flash-Next (self-host, Qwen Community License 1.0); this managed API is recorded proprietary/API-only. Self-reported benchmarks carry over from Flash-Next (SWE-bench Pro 62.5, DeepSWE 58.7) — all vendor numbers, unverified by independent labs at launch. Thinking on by default.

An open-weight, experimental preview of the architecture that will underpin Qwen4, released Aug 26 2026 (Qwen/Qwen3.8-Flash-Next). A sparse MoE with ~6B active parameters (headline 125B-with-6B-activated; ~180B stored once a 51B n-gram embedding table and 4B multi-token-prediction module are counted), 512 experts (10 routed + 1 shared), and a hybrid Gated-DeltaNet + Qwen Sparse Attention design. Native 262,144-token context, extensible to 1M via YaRN. Accepts text, image, and video in and returns text out. Distinct from the managed Qwen Cloud 'Qwen3.8-Flash' API (which defaults to 1M context and bundled tools); this Next build is catalog/self-host only with no hosted list price at launch, served via Transformers, vLLM, SGLang, and TokenSpeed. Weights under the Qwen Community License 1.0. Self-reported vs DeepSeek-V4-Flash-0731: DeepSWE 58.7 vs 54.4, SWE-bench Pro 62.5 vs 56.0, LiveCodeBench v6 91.9, GPQA Diamond 91.7, though NL2Repo 48.1 vs 54.2 is a regression; vision self-reports include AndroidWorld 84.5 and RealWorldQA 88.5 — all vendor numbers, unverified at launch. Thinking on by default.

Z.ai's first natively multimodal GLM-5 (text, image, and video understanding in one stack), released Aug 26 2026 and stealth-tested beforehand as 'ox-alpha'. A 320B-total / 18B-active MoE (45 layers, hybrid linear + sparse attention) with a 1M-token context window and MIT-licensed open weights (zai-org/GLM-5.3-Flash). It is a distinct model from the text-only flagship GLM-5.3 (753B, $1.40/$4.40) and is priced roughly 10x cheaper on input: list $0.15 / $0.03 cached / $0.50 per Mtok, with a 50% promo through 2026-09-09. Vision sits inside the coding/agent loop (self-visual judgment) rather than as a bolted-on VL head. Self-reported vs GLM-5.2: DeepSWE 63.4 vs 46.2, AutomationBench 48.8 vs 26.2, Terminal-Bench 2.1 84.3 vs 81.0; vision self-reports include CharXiv Reasoning 89.4 and Chartography 78.0, though BabyVision 53.4 trails Gemini 3.7 Flash — all vendor numbers, unverified at launch. Thinking is always on and cannot be disabled.

The smallest member of IBM's Granite 4.2 open reasoning family, released Aug 25 2026 under Apache-2.0 (ibm-granite/granite-4.2-3b; reports ~4B parameters on Hugging Face) and aimed at local / edge deployment. A dense, decoder-only transformer with the family's thinking / non-thinking switch and low-effort thinking mode, pre-trained from scratch on ~15T tokens with a five-phase curriculum (context extended to a claimed 512K; shipped configuration 131,072 tokens), then SFT on reasoning data and multi-stage RL. Supports native tool calling. Open weights on Hugging Face, Ollama, and GitHub; no hosted list price at launch.

The mid-size member of IBM's Granite 4.2 open reasoning family, released Aug 25 2026 under Apache-2.0 (ibm-granite/granite-4.2-8b; reports ~9B parameters on Hugging Face). A dense, decoder-only transformer sharing the family's thinking / non-thinking switch and low-effort thinking mode, pre-trained from scratch on ~15T tokens with a five-phase curriculum (context extended to a claimed 512K; shipped configuration 131,072 tokens), then SFT on reasoning/agentic-trajectory data and multi-stage RL. Like the 30B, it is trained to call tools and act inside real sandboxed environments for multi-step software-engineering, terminal, and search-driven tasks. Open weights on Hugging Face, Ollama, and GitHub; no hosted list price at launch.

The flagship of IBM's Granite 4.2 family, released Aug 25 2026 under Apache-2.0 (ibm-granite/granite-4.2-30b; reports ~29B parameters on Hugging Face). A dense, decoder-only transformer with a thinking / non-thinking switch so one checkpoint can either reason step by step or answer directly, plus a low-effort thinking mode that caps the reasoning budget on easy queries. Pre-trained from scratch on ~15T tokens with a five-phase curriculum that extends context to a claimed 512K (shipped configuration 131,072 tokens), then supervised fine-tuned on chain-of-thought / reasoning / agentic-trajectory data and post-trained with multi-stage RL. Aimed squarely at agent work — native tool calling, multi-step software engineering, terminal tasks, and search-driven workflows — with the 8B and 30B trained to act with tools inside real sandboxed environments. Open weights on Hugging Face, Ollama, and GitHub; no per-token hosted list price at launch.

The flagship model of Apodex's Aug 24 2026 release, built for what the lab calls "working capability" — sustained, verifiable progress on long-horizon professional and scientific tasks rather than one-shot answers. It powers an online workbench where a single long-running task can read files (papers, CSVs, spreadsheets, images, code), search, execute analysis code, revise its own plan, and coordinate an asynchronous "Agent Team" of sub-agents, with a separate "Statement Review" step that independently checks key claims before delivery. Trained with SFT plus agentic RL (the lab's PIVOT-RL method concentrates learning on the consequential decision points of long trajectories) across two scaling axes: Environment Scaling and Agentic Coordination Scaling. Apodex reports the full model entering the leading band of current agentic systems on complex professional work, finance, and scientific research while retaining broad reasoning, math, coding, and deep-search coverage — self-reported figures (e.g. APEX-Agents 38.5, GDPVal 78.8, FrontierFinance 54.3, FrontierScience-Research 63.3) with no independent audit at launch. Parameter count is not disclosed; the lab explicitly declines size comparisons. Available now via the apodex.ai web workbench with bare-model API access rolling out through major platforms; open weights and full developer docs are in progress.

Thomson Reuters' first in-house proprietary large language model, launched August 24 2026. Built by starting from a strong (undisclosed) open-source foundation and applying state-of-the-art mid- and post-training on decades of proprietary Thomson Reuters content — Westlaw, Practical Law, Checkpoint, and Reuters — with hundreds of subject-matter experts involved from training-objective design through final evaluation, to the company's 'Fiduciary-Grade' standard. Total training investment is reported at ~$40M (talent + compute), positioned as reaching frontier-level quality at a fraction of typical frontier cost while remaining fully owned and controlled by Thomson Reuters. The company claims early evaluations put Thomson 'on par with the latest frontier models' across a range of tasks, with notable domain-specific uplift in instruction following and dense professional-content reasoning — vendor claims, corroborated only by limited external academic testing at launch and with no public benchmark table, so treat as unverified. Model size, architecture, and context window are undisclosed. First deployment is inside Tabular Analysis in CoCounsel Legal (which remains multi-model by design); broader rollout across the legal and tax portfolio plus sovereign-AI options is planned. A smaller version is being released as an open-weight model on Hugging Face for academic and non-commercial use; the flagship remains proprietary and product/API-only.

DeepSeek's first multimodal V4 model — an experimental vision-understanding checkpoint that went live on the DeepSeek API (model='deepseek-v4-flash-vision-exp') on Aug 21, 2026. It extends DeepSeek-V4-Flash with image understanding while keeping its full text capabilities (agents, reasoning, coding, and world knowledge), matching V4-Flash on text benchmarks. DeepSeek reports a major jump on multimodal agent benchmarks over V4-Flash, bringing multimodal-agent performance close to Opus-4.8 — a vendor-reported result, unverified by an independent harness at launch. Keeps V4-Flash's 284B-total / 13B-active sparse MoE architecture and 1M-token context, with up to ~393K output tokens; accepts text plus up to 600 images per request (8,192px per side, 64 MiB payload) and returns text only, with images billed at up to 384 tokens each. API pricing held at V4-Flash rates: $0.22 / $0.66 per 1M input/output tokens, with a $0.007 per 1M cached-input rate. API-only and experimental at launch — weights were not published, so treated as proprietary. RETIRED 2026-09-10: superseded by DeepSeek-V4.1-Flash, whose native multimodal support absorbs this experiment. For compatibility the `deepseek-v4-flash-vision-exp` API id temporarily routes to V4.1-Flash.

The flagship of Tencent Hunyuan's Hy-MT2 family of 'fast-thinking' multilingual machine-translation models, open-weighted on Hugging Face on Aug 20 2026. A Mixture-of-Experts model with 30B total and ~3B active parameters covering 33 language pairs plus five Chinese-dialect and minority-language pairs, with workflows for structured/delimiter-based, contextual, glossary-based, and style-guided translation. Runs a short 8,192-token context with up to 4,096 output tokens and is small enough to run locally. Tencent reports it outperforming open heavyweights such as DeepSeek-V4-Pro and Kimi K2.6 on translation quality, with even the smaller 1.8B sibling (Hy-MT2-1.8B, released the same day) beating commercial APIs from Microsoft and Doubao — vendor figures, unverified at launch. A specialized translation LLM (text in/out), included as in-scope; the smaller 1.8B and FP8 variants are not tracked separately.

The flagship of DeepReinforce's Ornith-1.5 family, released 2026-08-19 under the MIT license with weights on Hugging Face. A ~397B-parameter Mixture-of-Experts coding/agent model (per-token active count not disclosed) trained with a self-improving RL loop: rather than fixed human-curated tasks, the system proposes progressively harder tasks itself, generates a task-specific orchestration scaffold for each, and produces the solution rollouts used for reinforcement learning, with reward propagating across all three stages (all optimized with GRPO). Vendor-reported, five-run-averaged figures: Terminal-Bench 2.1 85.1 and DeepSWE 56.0 — which DeepReinforce puts on par with Claude Opus 4.8 (85.0 / 59.0) and ahead of GLM-5.2 and DeepSeek-V4-Flash-0731 at comparable scale — plus 92.8 GPQA Diamond and 86.6 BrowseComp. All numbers are the vendor's own and unverified by an independent harness at launch. Extends the self-scaffolding approach introduced in Ornith-1.0 (June 2026).

The first open-weight release in the dots3 series from Dots Studio (rednote-hilab), the AI lab of Xiaohongshu (RedNote), open-weighted on Hugging Face on Aug 14 2026 under Apache-2.0. A Mixture-of-Experts model with ~280B total parameters and ~16B active per token, carrying a 512K-token context and multimodal understanding across text, vision, and audio (text output). Positioned less around benchmark-maxxing and more around combining reasoning, long-context processing, coding, and multi-step agent workflows in a compute-efficient architecture optimized for long-horizon real-world tasks. Introduces TEMPO, a reinforcement-learning method for long-horizon agents in which the model periodically checkpoints its own progress and updates its working memory mid-task. Comes from the same dots3 series whose internal build scored a perfect 42/42 at the 2026 International Mathematical Olympiad. Weights ship in BF16 and FP8; served free on OpenRouter (dots-studio/dots3-note-preview) at launch.

The dated GA build behind DeepSeek's 'deepseek-v4-pro' API id, pinned on the official pricing table with an OpenRouter listing dated Aug 12 2026 — the Pro-tier counterpart to the way V4-Flash graduated as DeepSeek-V4-Flash-0731. It is a quiet version pin (no separate launch post or benchmark card), keeping the V4-Pro architecture and the 1M-token context / ~384K max-output window. List pricing is cache-heavy: $0.435 input cache-miss / $0.003625 cache-hit / $0.87 output per Mtok, with concurrency 500 (vs Flash's 2500); DeepSeek warns a significant, undated API price increase is coming. Thinking is on by default at effort 'high' (requested medium/xhigh both collapse to high). Architecture figures (1.6T total / 49B active MoE, hybrid long-context attention) are carried over from the April 2026 V4-Pro preview and are not independently reconfirmed for the 0813 build; Hugging Face still hosts only the April preview weights (MIT), with no confirmed separate 0813 open-weight repo, so this row is recorded as proprietary/API-only. DEPRECATED 2026-09-14: the `deepseek-v4-pro` API id this dated build served now routes to DeepSeek-V4.1-Flash at Flash rates, pending the V4.1-Pro launch.

xAI's frontier update to Grok 4.5, released 2026-08-12. Built on the same reported 1.5-trillion-parameter V9 foundation as Grok 4.5, with gains coming from upgraded supervised fine-tuning and reinforcement learning rather than a new base. Supports a 500K-token context and configurable reasoning effort (low, medium, high default, xhigh). List pricing is $2 / $6 per 1M input/output tokens below 200K prompt tokens ($0.50 cached input), rising to $4 / $12 above 200K.

NVIDIA's efficient open Mixture-of-Experts model, released Aug 11 2026 for long-running agents. A hybrid Mamba-2 + MoE + Attention design with ~31.6B total and ~3.6B active parameters and a 1M-token context, shipped alongside the NeMo Switchyard model router. NVIDIA reports performance comparable to gpt-oss-120b at roughly a quarter of the total parameters, up to 4x the output speed of similar-sized models, and 10,000 tasks completed ~30% faster than Qwen3.6-35B at similar accuracy. Vendor-reported BF16 figures: SWE-bench Verified 51.56, GPQA Diamond 75.44, MMLU Pro 81.94, PinchBench 85.37 — self-reported and unverified by an independent harness at launch. Ships under the permissive OpenMDW-1.1 license with weights, training data, and recipes released, and is available on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as an NVIDIA NIM microservice.

Sakana AI's Japanese-specialized reasoning and agent model, opened as an API on Aug 11 2026. Rather than pretraining from scratch, Sakana builds on Moonshot's open-weight Kimi K2.6 and adds training for Japanese language and business contexts — the lab's 'sovereign AI by tuning other labs' strategy. Handles Japanese instruction following, business writing, math, coding, research, and multi-step agentic workflows; accepts text, images, and files such as PDFs and returns text, over a 262,144-token context with up to 65,536 output tokens, and ships native web-search and code-execution tools through a unified OpenAI-compatible API. Priced on OpenRouter at $0.95 / $4.00 per 1M input/output tokens. Not available in the EU/EEA, the UK, or Switzerland. Served via API; weights not separately published (base Kimi K2.6 is open-weight).

Microsoft's upgraded inference-efficient agentic coding model, released Aug 11 2026 and rolling out in production across GitHub Copilot (CLI and VS Code). A post-training-focused successor to MAI-Code-1-Flash, tuned through hundreds of thousands of reinforcement-learning environments built from real GitHub Copilot usage rather than raw scaling. Microsoft reports a +22% gain on Terminal-Bench 2.1 for CLI tasks and +15% on .NET workloads over 1.0, alongside 25% greater token efficiency (25% fewer tokens per task, streaming 25% faster) — at roughly a quarter of the cost of MAI-Code-1.0. Text/code only; size and context window undisclosed, weights not released.

Meta's first open-weight agentic model, released Aug 10, 2026 under an Apache 2.0 license — Meta's return to open weights after the closed Muse Spark line. A ~30B-parameter dense causal transformer (about 29.6B parameters across 52 layers) paired with a ~1.8B ViT-G/14 perception encoder, so it accepts interleaved text and images and returns text across more than 100 languages. Carries a 131,072-token context, a 202,048-token vocabulary, and a Jan 4, 2026 knowledge cutoff. Uses grouped-query attention (32 query heads, 2 KV heads) in a local/local/local/global pattern with a 2,048-token sliding window, plus speculative decoding for throughput. Quantized to roughly 4-bit it fits inside a ~24GB memory envelope, running on a single consumer GPU or an Apple-silicon Mac — the model is tuned for on-device agent workloads. Shipped the same week as the closed-weight Muse Spark 1.2 coding flagship; Meta's first agentic model to ship with both open weights and a permissive commercial-use license.

The flagship of South Korean lab Upstage's in-house Solar family, surfaced on developer platforms around Aug 10 2026 and formally unveiled Aug 14 2026. An agent-first reasoning model pitched as a reliable workhorse — 'save frontier models for frontier problems' — with strong office-productivity, document-intensive, and coding performance rather than raw frontier scores. Carries a 524K-token context and up to 131K output tokens. Upstage reports large gains over Solar Pro 3: ~71 on the AA-LCR long-document benchmark (about 2.3x its predecessor) and 23 on Tau3-Banking tool-use (a 2.6x jump); Artificial Analysis scores it 42 on its Intelligence Index. API pricing is $0.30 / $1.20 per 1M input/output tokens, with a launch promo cutting that 90% to $0.03 / $0.12. Upstage is the first Korean company registered as an official model provider alongside OpenAI, Anthropic, Google, and NVIDIA. Served via Upstage's API and OpenRouter; weights were not published at launch, so treated as proprietary.

Liquid AI's on-device agentic model, released Aug 6 2026 (surfaced on hosted platforms ~Aug 11) — a 2.69B-parameter dense model, distinct from the LFM2.5-8B-A1B MoE. Uses Liquid's hybrid stack across 30 layers: 22 double-gated short-convolution blocks plus 8 grouped-query-attention blocks, with a 128K-token vocabulary and a 131,072-token context, pre-trained on ~34T tokens across 16 languages. Text-only, built to plan, call tools, and complete multi-step tasks entirely on-device (phone, laptop, PC, or robot) so data never leaves the device and the marginal cost per run is near zero; Liquid reports tool-use and instruction-following competitive with models ~4x its size (e.g. leading Qwen3.5-9B on ToolSandbox, Multi-IF, and IFStruct), while explicitly not recommending it for agentic coding or knowledge-heavy work. Decodes at ~220 tok/s on an Apple M5 Max in under 2.5GB and ~30 tok/s on a phone. Open weights (base + post-trained) on Hugging Face under the LFM Open License (lfm1.0), shipped day-one in GGUF, MLX, and ONNX.

The smallest member of Ant Group inclusionAI's Ling 3.0 family, open-weighted on Hugging Face on Aug 6 2026 under the MIT license — distinct from the (API-only at launch) Ling-3.0-flash. A sparse Mixture-of-Experts model with 7.9B total parameters and only ~1.3B active per token: 128 routed experts with 8 routed plus 1 shared expert active per token, using the same 3:1 alternating stack of Kimi Delta Attention (KDA, linear) and Multi-head Latent Attention (MLA) layers as the rest of the family, for a 262,144-token (256K) context. Pitched as a highly economical on-device agent/reasoning model; weights are provided in BF16, FP8, and INT4 for a wide range of hardware. Vendor benchmark figures are unverified at launch.

Meta's flagship coding model, released 2026-08-05 and purpose-built for complex software engineering: debugging sprawling codebases, validating changes across thousands of files, and multi-step reasoning, with deep integration into persistent asynchronous background agents. Accepts text, image, video, audio, and PDF input over a 1M-token context and returns text. Standard-tier API pricing is $1.25 / $4.25 per 1M input/output tokens ($0.15 cached input); a new muse-spark-1.2-contributor tier drops to $0.10 / $0.20 in exchange for permission to train future Meta models on your prompts and completions. Shipped alongside Muse Code, a terminal-based coding agent powered by the model. Meta has signaled open weights are coming.

The production release of DeepSeek's V4-Flash tier — the April V4-Flash preview retrained on a substantially improved post-training pipeline targeting coding, agents, reasoning, and tool use, with no change to the base architecture. Retains 284B total / 13B active parameters (MoE) and the 1M-token context window. DeepSeek reports the 0731 build scoring higher than its own larger V4-Pro-Preview on all nine agent and coding benchmarks it published — a vendor-reported result, with independent replication still limited at launch. Weights released on Hugging Face under the MIT license; API pricing held at $0.14 / $0.28 per Mtok. The upgrade is silent for existing callers: same endpoint, same key, same deepseek-v4-flash model name, zero migration cost. RETIRED 2026-09-10: the `deepseek-v4-flash` API id this dated build served was retired in favour of DeepSeek-V4.1-Flash and now routes there.

Cost-optimized multimodal member of the Qwen3.7 line — a vision-language reasoning model with a 1M-token context, tuned for high-volume multimodal agent workloads (visual coding, screen perception, browser/computer use, search) where cost matters more than peak intelligence. Launched quietly on Jul 27, 2026 as an OpenRouter/API listing at $0.03/$0.13 per 1M tokens, making it the cheapest 1M-context multimodal model available at release. Closed-weights and API-only; Alibaba published no technical report, benchmark suite, or architecture details, though community speculation points to a small sparse-MoE design.

Anthropic's flagship Opus model, released July 24, 2026 — positioned as the go-to model for most knowledge work and automation, approaching the capability of Claude Fable 5 in many categories at roughly half the price. Built for demanding reasoning, autonomous coding, software development, and long-horizon agentic work. Introduces a five-level 'effort' dial exposed to developers on the Claude API and Platform, letting them trade compute and tokens for capability — at lower effort it preserves much of its performance while using fewer tokens and costing less to run. 1M-token context window (available at standard token pricing, not a separate long-context surcharge) with up to 128K output tokens; text, vision, and code. Standard API pricing $5/$25 per Mtok (the same as its predecessor Opus 4.8), plus a Fast mode at $10/$50. Anthropic describes it as its most aligned Opus model and the least susceptible to being tricked into misuse. Becomes the default model for Claude Max subscribers and is available across Anthropic's paid plans; scores 61 on the Artificial Analysis Intelligence Index. Closed weights; architecture, parameter count, and training compute undisclosed.

Ant Group's efficiency-focused Mixture-of-Experts model, released July 23 2026 by its inclusionAI lab: 124B total parameters activating only ~5.1B per token (1/64 expert activation). Ant claims it matches or beats the company's own ~1T-parameter Ling-2.6 flagship on most benchmarks it shows, at 1/8 the total and 1/12 the active parameters — a vendor claim with no public benchmark table or independent audit at launch, so treat it as unverified. Built for production-scale agents (MCP tool use, multi-agent coordination) rather than chat, with both thinking and non-thinking modes. Architecture is a native hybrid-linear attention stack interleaving Kimi Delta Attention (KDA) and Multi-head Latent Attention (MLA) at a reported 5:1 ratio, giving an economical 262,144-token (256K) context, with 1M cited as the scaling target. Ant docs claim peak inference up to 1,000 tokens/s and <100ms time-to-first-token on its own stack. Announced as open-weight under Apache 2.0, but as of July 24 no weights or model card were posted to the inclusionAI Hugging Face org — so the license and open-weight status are announced but unconfirmed (weights not yet downloadable; not self-hostable today). Usable now only via hosted API — free on OpenRouter (as inclusionai/ling-3.0-flash:free, hosted by Novita) and Vercel's AI Gateway through August 3 2026; no post-promo per-token price published at launch.

Google DeepMind's fastest and most cost-effective 3.5-class model, released July 21, 2026 for low-latency and high-throughput agentic workloads like agentic search and document processing. Runs at ~350 output tokens/s (Artificial Analysis) with configurable thinking levels and built-in computer use, priced at $0.30 / $2.50 per 1M input/output tokens. Multimodal over a 1M-token context and a large step up on 3.1 Flash-Lite: Terminal-Bench 2.1 54% (vs 31%), GDM-MRCR v2 72.2% (vs 60.1%), GDPval-AA v2 1140 (vs 642); on several agentic and coding evals it even surpasses 3 Flash (SWE-Bench Pro 54.2% vs 49.6%, OSWorld-Verified 74.0% vs 65.1%). Available in the Gemini API (AI Studio, Android Studio), Gemini Enterprise, the Gemini app, and rolling out in Google Search.

Google DeepMind's July 2026 workhorse Flash model, built for scaling agentic workflows. Multimodal over a 1M-token context, it improves on Gemini 3.5 Flash in coding, knowledge work, and computer use while cutting output-token usage ~17% (up to 65% on some benchmarks like DeepSWE) and taking fewer reasoning steps and tool calls. Ships at a lower price than 3.5 Flash ($1.50 / $7.50 per 1M input/output tokens). Google-reported gains: DeepSWE 49% (vs 37%), MLE-Bench 63.9% (vs 49.7%), OSWorld-Verified 83.0% (vs 78.4%), GDPval-AA v2 1421 (vs 1349); knowledge cutoff advances to March 2026. Computer use is a built-in client-side tool. Available in the Gemini API (AI Studio, Android Studio, Antigravity), Gemini Enterprise, and the Gemini app.

Poolside's open-weight agentic-coding model and a scale-up of the Laguna XS family (same pre-training data as XS 2.1): a 118B-total / 8B-active Mixture-of-Experts that activates only ~6.8% of its parameters per token, giving larger-model behavior while staying cheap to serve, with a 1M-token context in both thinking and no-thinking modes. Pitched by Poolside as 'the West's most capable open-weight model' — the claim is about its weight class, not the outright frontier. Two modes (off / max, max default; the model sets its own test-time compute budget). Vendor-reported: Terminal-Bench 2.1 70.2% and SWE-bench Multilingual 78.5% (tops the published open disclosed-size table), plus SWE-bench Pro 59.4%, DeepSWE v1.1 40.4%, SWE Atlas 46.2%, Toolathlon Verified 49.7% — matching or beating models several times its size, though closed frontier models still lead outright. Trained in under nine weeks on 4,096 NVIDIA H200 GPUs (pre-training began 22 May 2026); first Poolside model with RL in FP8. Knowledge cutoff November 2025. Weights on Hugging Face under the permissive OpenMDW-1.1 license in BF16/FP8/INT4/NVFP4 with GGUF/MLX conversions and DFlash draft models; at 4-bit it runs on a single NVIDIA DGX Spark. Day-one support for vLLM, SGLang, and Ollama; hosted free at 256K context via OpenRouter and paid at the full 1M context ($0.10 / $0.20 / $0.01 per 1M input / output / cache-read tokens), also on Baseten, Kilo, Prime Intellect, and ZML.

Moonshot's flagship open-weight agentic model and the largest open model released to date: a 2.8T-parameter MoE (896 experts, 16 active per token) using Kimi Delta Attention and Attention Residuals, with native multimodal input and a 1M-token context. Launched via API on Jul 16, 2026 at $3/$15 per Mtok (cached input $0.30); full open weights published to Hugging Face on Jul 26, 2026 — a day ahead of the announced Jul 27 target — under a Modified MIT license, making it freely downloadable and self-hostable.

Thinking Machines Lab's first model and the leading U.S. open-weights release: a natively multimodal Mixture-of-Experts with 975B total / 41B active parameters that reasons across text, image, and audio inputs and emits text. Pretrained on ~45T tokens; served with a 1M-token context from the Hugging Face weights (256K on the hosted Tinker API). Apache-2.0 licensed (BF16 + NVFP4 checkpoints on Hugging Face), built for developers fine-tuning on proprietary data — coding assistants, agents/tool use, chatbots, and RAG — with an explicit low-cost and censorship-resistance focus. Debuted at 41 on the Artificial Analysis Intelligence Index. Hosted pricing (256K) $3.74/$9.36 per Mtok reflects a limited-time 50% launch discount.

The efficient, ~32B variant of Kwaipilot's KAT-Coder V2.5, optimized through multi-stage training (supervised fine-tuning plus reinforcement learning). Shares KAT-Coder-Pro's 256K-token context and 80K max output, with function calling and tool use for agentic coding, at roughly a fifth of Pro's cost: $0.15 / $0.60 per million input/output tokens. Surfaced on release trackers on July 14 2026.

Kwaipilot's flagship agentic coding model, from the KAT (Kwaipilot Agentic Tuning) series at Kuaishou. A Mixture-of-Experts model with ~72B active parameters, trained through large-scale agentic reinforcement learning in reconstructed, verifiable repository environments. Supports function calling, tool use, structured JSON output, and prompt caching, with a 256K-token context and up to 80K output tokens. Served via API (StreamLake / Atlas Cloud / OpenRouter) at $0.74 / $2.96 per million input/output tokens. Surfaced on release trackers on July 14 2026.

OpenAI's GPT-5.6 series umbrella row. Officially previewed June 26, 2026 as three durable capability tiers — Sol (flagship), Terra (balanced, for everyday work), and Luna (fast and affordable) — introduced with a new `max` reasoning effort for deeper reasoning and an `ultra` mode that leverages subagents to accelerate complex work. In the GPT-5.6 naming system the number marks the generation while Sol/Terra/Luna are tiers that can advance on their own cadence. Initially a limited preview via the API and Codex for a small group of vetted partners (after U.S. government review), with general availability across ChatGPT, Codex, and the API planned in the following weeks.

Balanced, everyday-work tier of OpenAI's GPT-5.6 series, officially previewed June 26, 2026. OpenAI positions Terra as competitive with GPT-5.5 while being roughly 2x cheaper. Shares the series' new `max` reasoning effort and `ultra` subagent mode and OpenAI's GPT-5.6 safety stack. Begins as a limited preview via the API and Codex for vetted partners after U.S. government review, with broad availability planned in the following weeks. Priced at $2.50 / $15 per million input/output tokens.

Flagship tier of OpenAI's GPT-5.6 series, officially previewed June 26, 2026 — OpenAI's strongest model to date. Adds a new `max` reasoning effort for the deepest reasoning and an `ultra` mode that uses subagents to accelerate complex work. Sets a new state of the art on Terminal-Bench 2.1 (command-line, agentic coding) and shows broad gains in long-horizon biology (GeneBench v1) and cybersecurity (ExploitBench, ExploitGym), paired with OpenAI's most robust safety stack and a phased release. Begins as a limited preview via the API and Codex for a small group of vetted partners after U.S. government review, with general availability planned in the following weeks; also launching on Cerebras at up to 750 tokens/sec in July. Priced at $5 / $30 per million input/output tokens.

Meta Superintelligence Labs' first paid model, released July 9, 2026 in US public preview on the Meta Model API. A natively multimodal reasoning model (text, image, video, PDF, and audio input; text output) with explicit chain-of-thought reasoning and a 1M-token context window that the model actively compacts. Positioned for agentic work and coding — tool use, multi-step workflow coordination, and long-horizon autonomous tasks — and pitched by Meta as roughly a quarter of the price of comparable Anthropic and OpenAI models at $1.25 in / $4.25 out per Mtok (with $20 in free credits per new API account). Marks the first time Meta has charged businesses for one of its models, a departure from the open-weight Llama strategy. Closed weights, undisclosed size. Vendor and third-party benchmarks place it around the Opus 4.8 / GPT-5.5 tier — strongest as an agent/workflow model and in tool-augmented reasoning, competitive but not dominant on coding and multimodal tasks.

SpaceXAI's "Opus-class" agentic flagship, released July 8, 2026 — the first Grok model trained jointly with the coding startup Cursor (Anysphere), on trillions of tokens of real Cursor usage data plus STEM tasks, research papers, and other knowledge work. Targets software engineering, agentic tasks, and knowledge work, with explicit strength in legal and finance use cases (SpaceXAI claims the top spot on the Harvey Legal Agent Benchmark). A Mixture-of-Experts model reported to be built on a ~1.5-trillion-parameter "V9" foundation. Vendor benchmarks are mixed against Anthropic's Opus 4.8 — ahead on DeepSWE 1.0 and Terminal-Bench 2.1, behind on DeepSWE 1.1 and SWE-bench Pro — but markedly more token-efficient (~15,900 output tokens on SWE-bench Pro tasks, ~4.2x fewer than Opus 4.8) and served at ~80 tokens/sec. Base pricing $2/$6 per Mtok; Cursor lists a faster variant at $4/$18. Available in Grok Build (default model), Cursor (all plans), and the SpaceXAI console; initially unavailable in the EU (expected mid-July). xAI was absorbed by SpaceX in Feb 2026 and rebranded SpaceXAI.

The general-availability release of Tencent's third-generation Hunyuan (Hunyuan 3.0), officially launched and open-sourced on July 6, 2026 after April's "Hy3 preview". A 295B-total / 21B-active Transformer MoE with an additional 3.8B multi-token-prediction (MTP) layer and a 256K-token context, offering three selectable inference modes that blend fast and slow thinking. Positioned as a leading open model for its size and cost efficiency, with standout results in coding, search, and scientific reasoning: Tencent reports it rivals flagship open models such as GLM-5.2 and DeepSeek-V4 (at 2-5x the active parameters) and matches or surpasses GPT-5.5 on several science benchmarks. Vendor-reported scores include 78.0 on SWE-bench Verified and 57.9 on SWE-bench Pro. Now Apache-2.0 licensed (the preview used Tencent's community license), with weights on Hugging Face (tencent/Hy3) and ModelScope and a free two-week API route on OpenRouter (tencent/hy3:free) through July 21, 2026. Deeply integrated into WeChat and Tencent's core products.

A deployment-optimized open-weight model from NVIDIA, released July 6, 2026 — a compressed variant of Nemotron-3-Super-120B-A12B produced with "Iterative Puzzle", a post-training compression framework that jointly prunes MoE experts, active-parameter budget, and Mamba state to boost inference efficiency while preserving accuracy. Reduces the parent from 120.7B total / 12.8B active to 75.3B total / 9.3B active, keeping the hybrid Mamba-Transformer LatentMoE architecture with Multi-Token Prediction. Delivers ~2x higher server throughput than Nemotron-3-Super on a single 8xB200 node at matched user throughput and raises sustainable 1M-token single-H100 concurrency from 1 to 8 requests. Targets collaborative agents, chatbots, RAG, complex instruction-following, and long-context reasoning across English, code, and six other languages. Shipped in BF16, FP8, and NVFP4 variants under the OpenMDW-1.1 license.

Rumored / early-access: Mistral AI has confirmed a new open-weight Mixture-of-Experts family — described by CEO Arthur Mensch as "fat but sparse" — aimed at closing the gap with frontier open-weight releases, with early access beginning in July 2026. Mensch confirmed the intent but disclosed almost nothing concrete: no parameter count, no benchmarks, no license terms, and no release date. Tracked as rumored until weights or an official product page land.

Poolside's open-weight small coding model: a 33B-total / 3B-active Mixture-of-Experts built for agentic coding and long-horizon work on a local machine, served at 256K context. An upgraded XS.2 (same architecture) that lifts SWE-bench Multilingual by 5.4 points to 63.1% and improves terminal-style tasks. Ships with open-weighted DFlash speculator (draft) models for each checkpoint that roughly double local tokens/sec, plus BF16/FP8/INT4/NVFP4 quantized checkpoints; supported in vLLM, SGLang, TensorRT-LLM, HF transformers, and Ollama (llama.cpp coming). Newly relicensed under the fully permissive OpenMDW-1.1. Available free on Hugging Face and via a free OpenRouter tier, with paid API pricing of $0.10 / $0.20 / $0.05 per 1M input / output / cache-read tokens. Its predecessor Laguna XS.2 sunsets on Poolside's API one week after launch.

Anthropic's most agentic Sonnet model yet, with performance approaching Claude Opus 4.8 at a lower price. Built for coding, tool use (browsers and terminals), and autonomous multi-step agentic work, with selectable effort levels up to xhigh. A substantial upgrade over its predecessor Sonnet 4.6 on reasoning, tool use, coding, and knowledge work, with gains shown on SWE-bench, OSWorld-Verified, BrowseComp, and Humanity's Last Exam. The default model on the Free and Pro plans and available to Max, Team, and Enterprise users; in Claude Code and via the Claude API as claude-sonnet-5. Uses an updated tokenizer (same approach as Opus 4.7). Ships with real-time cyber safeguards enabled by default. Introductory API pricing of $2/$10 per Mtok through Aug 31, 2026, then standard $3/$15.

Meituan's open-weight flagship: a 1.6-trillion-parameter Mixture-of-Experts model (~48B active per token, dynamically routed between ~33B and ~56B) with a 1M-token context, built for agentic coding. Notable as the largest Chinese model trained — for both pre-training and inference — entirely on a ~50,000-card cluster of domestic Chinese AI chips (Meituan's use of the Huawei Collective Communication Library points to Huawei Ascend hardware), and the first trillion-parameter model Meituan claims completed full-process training on home-grown compute. Vendor-reported software-engineering results: 59.5 on SWE-bench Pro (ahead of GPT-5.5's 58.6), 70.8 on Terminal-Bench 2.1, and 77.3 on SWE-bench Multilingual, with overall quality positioned as comparable to Gemini 3.1 Pro (self-reported, not yet independently verified). Open-sourced under the MIT license with weights on Hugging Face and GitHub; follows LongCat-Flash (560B, Sep 2025) and the multimodal LongCat-Next (Mar 2026).

ByteDance's flagship next-generation agent model (served as Doubao-Seed-2.1-pro), built for the "coding and agent era." A deep-thinking model tuned for strong demand understanding, long-horizon planning, and continuous self-repair across complex coding, long-chain agents, and multi-step engineering delivery, with a 256K-token context. ByteDance reports its core coding, agent, and multimodal capabilities are comparable to GPT-5.5, with the highest score on GDPVal, top-tier results on the Agents' Last Exam, the highest score on MobileWorld, and SOTA results across several visual and video-understanding benchmarks (CharXiv-RQ, MeasureBench, TVBench, TOMATO). Proprietary; available via Doubao and Volcano Engine. List price ¥6 / ¥30 per million input/output tokens.

Sakana AI's frontier-class orchestration model: a single ~7B language model trained to coordinate a swappable pool of external frontier LLMs (model selection, delegation, verification, and synthesis happen internally), exposed behind one OpenAI-compatible API. The Ultra tier is tuned for maximum answer quality on hard, multi-step problems and coordinates a deeper, fixed pool of expert agents. Sakana reports it stands shoulder-to-shoulder with Anthropic's Fable 5 and Mythos Preview across coding, science, and reasoning benchmarks while routing around single-vendor/export-control risk. Current model ID fugu-ultra-20260615; vendor-reported scores include 95.5 GPQA-D, 73.7 SWE-Bench Pro, 93.2 LiveCodeBench, and 50.0 Humanity's Last Exam.

Sakana AI's orchestration model: a single language model that delivers a full multi-agent system behind one OpenAI-compatible API, dynamically routing tasks across a swappable pool of frontier LLMs (including recursive calls to itself). The base tier balances strong performance with low latency as an everyday default for coding, code review, and chat, and lets teams opt specific agents out of the pool for data/privacy/compliance needs. Released alongside Fugu Ultra on June 22, 2026; vendor-reported scores include 95.5 GPQA-D, 92.9 LiveCodeBench, and category-leading SciCode and long-context results.

Moonshot's open coding-focused agentic model built on K2.6, with native vision/video input, forced thinking mode, and stronger long-horizon software-engineering performance.

Z.ai's latest open flagship for long-horizon coding, agentic engineering, and million-token workflows, adding IndexShare sparse-attention reuse over GLM-5.1.

Native multimodal MiniMax model with a one-million-token context, sparse attention, and agentic coding/cowork positioning.

The public, guardrailed sibling of Claude Mythos 5 and Anthropic's most capable widely released model, built for long-horizon agentic work, coding, vision, and knowledge workflows. Launched June 9, 2026 across the Claude API, AWS, and Microsoft Foundry, then suspended three days later under a U.S. export-control directive. Anthropic says those controls were lifted June 30 and Fable 5 was restored globally on July 1 across Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with cloud partner access being re-enabled. Its safeguards route flagged cybersecurity, biology/chemistry, and distillation requests to Claude Opus 4.8.

Cohere's first developer-focused model and the first in its North family of code agents. A 30B-total / 3B-active MoE for agentic coding with a 256K context and up to 64K output, sized to run locally for enterprise coding agents. Apache-2.0.

Unisound's new-generation, general-purpose "native agentic" large model, built for task execution: it can autonomously decompose and advance complex real-world workflows of 100+ steps rather than single-turn Q&A. Unisound frames it around "high intelligence density x high token value" and reports ~25% lower thinking-token consumption. Available via the Unisound Token Hub.

NVIDIA's largest Nemotron 3 open-weight hybrid Mamba-Transformer MoE, tuned for agentic reasoning, coding, planning, and tool calling.

Microsoft's first in-house frontier reasoning model, unveiled at Build 2026. A sparse MoE (~35B active, 256K context) trained entirely on commercially licensed data without third-party distillation. Microsoft reports 97.0% on AIME 2025 and coding parity with Claude Opus 4.6 on SWE-bench Pro.

An inference-efficient agentic coding model from Microsoft (~5B active parameters), trained from the ground up on clean, traceable, enterprise-grade data without third-party distillation. Rolling out in GitHub Copilot; Microsoft reports a +16-point SWE-bench Pro lead over Claude Haiku 4.5.

Nex AGI's open-weight agentic flagship, post-trained on Qwen3.5-397B-A17B (397B total / ~17B active MoE) by the Shanghai Innovation Institute-led Nex alliance. Built around an "Agentic Thinking" framework for long-horizon coding, deep research, tool calling, and terminal execution; accepts text and image input and emits text with explicit reasoning traces and function calling. Apache-2.0, ~262K context. Nex reports parity with GPT-5.5 and Claude Opus 4.7 on several agentic and coding evals. A smaller Nex-N2-mini (35B/3B-active) was announced but is not yet open-sourced.

StepFun's high-efficiency multimodal sparse-MoE successor to Step-3.5-Flash: a ~196B-total / ~11B-active vision-language model with native image and video understanding, a 256K context, and selectable reasoning tiers (high/medium/low). Tuned for coding agents and search workflows.

Anthropic's most capable model, with strengthened agentic and long-running task performance.

Liquid AI's on-device Mixture-of-Experts model: 8.3B total parameters with only ~1.5B active per forward pass (32 experts, 4 active per token). Uses Liquid's hybrid architecture — 18 double-gated LIV convolution blocks plus 6 grouped-query-attention layers — for a 131K-token context that runs in under ~6GB of memory on consumer hardware. A reasoning-only model that emits an explicit chain of thought before its answer, with strong tool-calling and agentic performance for its size. Builds on the October 2025 LFM2-8B-A1B, expanding the context window to 128K and scaling pretraining from 12T to 38T tokens. Released May 28 2026 under the LFM Open License; caught in a July catalog-gap sweep.

Open-weight agentic model from MiniMax focused on real-world software engineering, office tasks, tool use, and self-improving training workflows.

Alibaba's proprietary flagship in the Qwen3.7 "Agent Frontier" line — a text-only sparse-MoE model with a 1M-token context, tuned for long-horizon agentic, coding, and reasoning workloads. Parameter count is undisclosed; access is API-only via Alibaba Cloud Model Studio / DashScope (and aggregators such as OpenRouter).

Announced at Google I/O 2026; emphasizes deep multimodal reasoning over a 2M-token context. Recent reporting says the broad launch slipped from June toward July while testers continue using it in Google Antigravity and LMArena.

Google's fast, cost-efficient Gemini 3.5 tier, unveiled at I/O 2026. Multimodal over a 1M-token context and tuned for agentic and coding workflows; Google says it beats Gemini 3.1 Pro on coding and tool-use while running ~4x faster.

Dense 27B that punches far above its weight on agentic coding — easy to self-host on a single GPU node.

Baidu's flagship ERNIE 5.1, derived from ERNIE 5.0 by extracting an optimal sub-network from its elastic sub-model matrix — compressing total parameters to ~1/3 and active parameters to ~1/2 of ERNIE 5.0 while reaching leading performance at only ~6% of the pre-training compute of comparable models. A Mixture-of-Experts model trained with a disaggregated fully-asynchronous RL stack and a multi-teacher on-policy-distillation pipeline, tuned for agentic execution, reasoning, world knowledge, and creative writing. Vendor-reported results: 99.6 on AIME26 (with tools, second only to Gemini 3.1 Pro), GPQA and MMLU-Pro approaching leading closed models, surpassing DeepSeek-V4-Pro on τ³-bench and SpreadsheetBench-Verified, and ranking 1st among Chinese models / 4th globally (score 1223) on the LMArena Search Arena. Proprietary; served via ERNIE Bot, Baidu AI Studio, and the Qianfan platform.

OpenAI's May 2026 GPT-5.5 release: a stronger frontier workhorse positioned for deep reasoning, coding, multimodal analysis, and long-context agent workflows. OpenAI lists an 800K-token input context and 128K-token output limit, with API pricing at $3 / $20 per million input/output tokens.

xAI's agentic flagship with a 1M-token context and aggressive API pricing.

Tencent's third-generation Hunyuan, rebuilt from scratch in ~90 days and open-sourced as the "Hy3 preview". A 295B-total / 21B-active Transformer MoE (80 layers, 192 experts with top-8 routing, plus a 3.8B multi-token-prediction layer) with a 256K-token context, positioned as a leading open reasoning-and-agent model for its size with strong cost efficiency. Vendor-reported results: 74.4 on SWE-bench Verified, 54.4 on Terminal-Bench 2.0, and 70.2 on WideSearch, with strong STEM-olympiad performance. Open weights on GitHub and Hugging Face under Tencent's community license.

Tencent Hunyuan open-weight fine-grained MoE model with 80B total parameters and 13B active parameters, optimized for agentic tool use.

Xiaomi's open-weight flagship: a 1.02T-parameter Mixture-of-Experts model with ~42B active parameters, a hybrid-attention architecture, and a 1M-token context window. Tuned for frontier-class agentic coding and long-horizon tasks (sustaining 1000+ tool calls with a proper harness). Open-sourced under the MIT license with weights and tokenizer on Hugging Face.

OpenAI's frontier reasoning model for life sciences, named after Rosalind Franklin and built to accelerate drug discovery, genomics, protein reasoning, and scientific research workflows. Optimized for multi-step, tool-heavy tasks (literature review, experimental design, sequence-to-function interpretation) with access to 50+ scientific databases via a Codex Life Sciences plugin. A June 3, 2026 update folded in GPT-5.5's agentic coding and tool use while using ~31% fewer tokens. Available as a research preview in ChatGPT, Codex, and the API through OpenAI's trusted-access program; not openly available.

Z.ai agentic-engineering follow-up to GLM-5, with stronger coding performance and better long-horizon tool-use behavior.

Google DeepMind's Gemma 4 advanced-reasoning open model for personal computers, part of the April 2026 Gemma 4 family.

Moonshot's open native multimodal agentic model for long-horizon coding, visual interface generation, and autonomous tool orchestration.

Mistral's March 2026 Small release: the first Mistral model to unify reasoning (Magistral), multimodal understanding (Pixtral), and agentic coding (Devstral) into one Apache 2.0 model. A 119B-total / ~6B-active Mixture-of-Experts (128 experts, 4 active per token) with native text+image input, a 256K context, and a configurable reasoning_effort toggle for fast or deep responses. API pricing is $0.15 / $0.60 per million input/output tokens.

StepFun's Apache-licensed sparse MoE model for fast agentic execution, coding, math, browsing, and tool-use workflows.

Apache-licensed Indian-context MoE from Sarvam AI, optimized for reasoning, coding, agentic tasks, and 22 Indian languages.

Workhorse GPT-5 release with a dedicated Thinking mode; widely deployed across ChatGPT and the API.

The flagship of Alibaba's small dense Qwen3.5 models. Independent analysis (Artificial Analysis) rated it the most intelligent model under 10B parameters at launch — roughly double the score of the next-closest sub-10B models — and the most intelligent multimodal model under 15B, leading peers on MMMU-Pro (~69%). A dense 9B with native vision, a 262K-token context, and the Qwen3.5 family's unified hybrid thinking / non-thinking mode. Native weights are BF16; in 4-bit it needs ~6GB, within reach of consumer laptops. High intelligence comes with heavy reasoning token usage (~260M output tokens to run the Intelligence Index).

A dense 4B in Alibaba's small Qwen3.5 family, rated by Artificial Analysis as the most intelligent model under 5B parameters at launch — outscoring several 7B–9B peers despite roughly half the parameters. Native vision, a 262K-token context, and the family's hybrid thinking / non-thinking mode; Apache-2.0 licensed. Scores ~65% on MMMU-Pro multimodal reasoning and runs in ~3GB at 4-bit, suitable for lightweight on-device agents.

A dense 2B Qwen3.5 model built for high-throughput, low-latency edge and on-device use. Despite its size it matches a 7B-class peer on Artificial Analysis's Intelligence Index. Apache-2.0, with native vision, a 262K-token context, and the family's hybrid thinking / non-thinking mode; runs in under 2GB at 4-bit, fitting laptops and smartphones.

The smallest Qwen3.5 model — a dense 0.8B designed for the most constrained on-device deployments, operating in non-thinking (instruct) mode by default. Apache-2.0, with native vision, a 262K-token context, and the family's hybrid thinking / non-thinking mode; needs roughly 2GB of VRAM and runs under 2GB at 4-bit, targeting smartphones and embedded hardware. Notable for a sub-1B model, it still scores ~26% on MMMU-Pro multimodal reasoning.

Z.ai flagship for complex systems engineering and long-horizon agentic tasks, scaling the GLM line to 744B total / 40B active parameters.

OpenAI's February 2026 Codex update, optimized for agentic software engineering in ChatGPT, Codex, and the API. GPT-5.3-Codex improves code quality, patch reliability, repository-scale reasoning, and long-running autonomous coding workflows while keeping the 400K-token input context and 128K-token output limit of the Codex line.

Open multimodal Kimi model that adds native visual agentic intelligence, instant and thinking modes, and agent-swarm workflows on top of the K2 base.

Coding-focused GLM release with improved multilingual agentic coding, terminal tasks, tool use, and interface generation.

OpenAI's December 2025 Codex model for agentic coding, released after GPT-5.2 with stronger repository understanding, code generation, and tool-use behavior for software engineering agents. The API model is listed with a 400K-token input context, 128K-token output limit, and $1.25 / $10 per million input/output tokens.

Ai2's fully open thinking model with public weights, code, data, checkpoints, and training details across the OLMo 3 pipeline.

Efficient Nemotron 3 MoE checkpoint for agentic reasoning and coding, activating about 3B parameters while supporting 1M-token contexts.

OpenAI's December 2025 GPT-5.2 general model release, positioned as a stronger default for reasoning, coding, vision, instruction following, and long-context analysis. OpenAI lists a 400K-token input context, 128K-token output limit, and $2 / $12 per million input/output tokens.

Mistral's largest open-weight MoE, aimed at frontier reasoning while remaining self-hostable.

Reasoning-first agent model that adds DeepSeek Sparse Attention and thinking directly inside tool-use workflows.

High-compute reasoning variant of V3.2, positioned for olympiad-level math, programming, and other deep reasoning tasks.

Open K2 reasoning-agent variant that interleaves step-by-step thinking with tool calls and supports stable 200-300 step tool-use trajectories.

Agentic reasoning and coding upgrade over GLM-4.5, expanding the text context window from 128K to 200K tokens.

September 2025 K2 update with stronger agentic coding, better frontend generation, and a doubled 256K context window.

Hybrid thinking/non-thinking release that upgraded tool calling, long-context training, and agent task performance.

ByteDance Seed's Apache-licensed long-context reasoning and agent model, with controllable thinking budgets and a native 512K context.

Smaller gpt-oss reasoning model optimized for local inference on systems with about 16GB of memory.

OpenAI's larger open-weight reasoning model, a 117B-total / 5.1B-active MoE with 128K context for local and self-hosted deployment.

Anthropic's August 2025 Opus point release, focused on stronger coding, reasoning, and agentic reliability over Claude Opus 4. Proprietary API model with text, vision, and code capabilities.

Google's enhanced Gemini 2.5 reasoning mode for harder math, science, coding, and multimodal analysis. Previewed at Google I/O 2025 and later made available to Gemini app subscribers, Deep Think uses more deliberative reasoning for complex prompts.

Open agentic, reasoning, and coding foundation model that marked Z.ai international rebrand and MIT-licensed GLM push.

Compact GLM-4.5 companion with 106B total / 12B active parameters for efficient agentic reasoning and coding.

LG AI Research's unified model with non-reasoning and reasoning modes, agentic tool use, and English, Korean, and Spanish support.

Original open K2 post-trained model: a 1T-parameter MoE optimized for coding, reasoning, and tool-using agentic workflows.

Hugging Face's fully open 3B multilingual long-context model with optional reasoning mode and 128K context.

Baidu's largest ERNIE 4.5 vision-language MoE, supporting text, image, and video inputs with thinking and non-thinking modes.

Updated MIT-licensed Kimi-VL reasoning model with better multimodal reasoning, video understanding, high-resolution perception, and lower thinking-token use.

Google's faster, lower-cost Gemini 2.5 model for high-throughput multimodal and agentic workloads. It brought Gemini 2.5's reasoning improvements to a production Flash tier with a 1M-token context and broad text, image, audio, video, and coding support.

Open Apache-licensed hybrid-attention reasoning model with 456B total / 45.9B active parameters and a native 1M-token context.

Mistral's first dedicated reasoning model family, released in Small open-weight and Medium enterprise/API tiers.

Open-weight 24B reasoning model from Mistral's Magistral family, popular for local reasoning experiments.

Major R1 reasoning update with stronger math, programming, general logic, function calling, and reduced hallucinations.

First Claude 4 Opus model, positioned for long-running agentic and coding work before the 4.x point releases.

ByteDance Seed reasoning model focused on long-horizon thinking and problem solving.

Sarvam's medium-scale open model for multilingual Indian-language chat, reasoning, and translation tasks.

Mistral and All Hands AI's open coding-agent model, released as a 24B Apache-2.0 research preview for software engineering tasks. Devstral is optimized for repository navigation, issue resolution, and agentic coding and is available via Hugging Face and Mistral's API.

Phi-4 reasoning-specialized model family for math, science, and chain-of-thought style tasks.

Largest open Qwen3 MoE, introducing hybrid thinking/non-thinking modes and 119-language coverage.

Reasoning model released alongside o4-mini with tool use, image reasoning, and stronger agentic problem solving.

Open Llama Nemotron reasoning model from NVIDIA's 2025 Nemotron family.

Post-R1 V3 update with improved reasoning, front-end coding, Chinese writing, search, and function calling.

Reasoning-focused Gemini 2.5 model that made thinking a core part of Google's flagship model line.

Baidu's reasoning model released alongside ERNIE 4.5 before the open ERNIE 4.5 weights.

Granite 3.2 update with reasoning controls and multimodal/document-oriented Granite variants.

Anthropic's first hybrid-reasoning Sonnet. Shut down May 11, 2026 as the 4.x line matured.

Nous reasoning-oriented Hermes model trained to combine concise answers with optional deep reasoning traces.

xAI's third-generation model family, introduced with stronger reasoning, search, and coding modes.

Popular local assistant model tuned for coding, math, function calling, and agentic workflows.

Vision-language Qwen2.5 model for image, document, video, and agentic visual grounding tasks.

Doubao 1.5 Pro update positioned for stronger multimodal, reasoning, and agentic work in Volcano Engine.

Breakout open reasoning model trained with large-scale reinforcement learning and released with weights under MIT.

Moonshot's multimodal reinforcement-learning reasoning model, reported as matching OpenAI o1 on math, coding, and multimodal reasoning.

Second-generation StepFun foundation model line with larger-scale multimodal and reasoning ambitions.

A 14B dense model that rivals far larger ones on math and reasoning, under a permissive MIT license.

First Gemini 2.0 release, built for native multimodal input/output, tool use, and agentic product integrations.

EXAONE 3.5 32B open-weight model for bilingual reasoning, coding, and long-context tasks.

General release of OpenAI's o1 reasoning model with stronger deliberative reasoning and multimodal ChatGPT integration.

Qwen's first public reasoning-preview model, aimed at math, coding, and deliberate problem solving.

Reasoning-preview model exposed in DeepSeek Chat ahead of the open DeepSeek-R1 release.

01.AI's MoE API model that reached the global top-10 on Chatbot Arena, strong in Chinese, math, and coding.

Broad Qwen2.5 foundation-model update spanning general, coding, math, and multimodal descendants.

OpenAI's first public reasoning-model preview, optimized to spend more inference time on hard math, coding, and science tasks.

Second-generation Grok release with Grok-2 and Grok-2 mini for chat, coding, reasoning, and image-enabled product experiences.

Major Sonnet upgrade that became Anthropic's default high-intelligence workhorse for coding, writing, and visual reasoning.

Qwen2's largest dense model, introducing stronger multilingual support, coding/math gains, and long-context variants.

Yi 1.5 update with stronger instruction following, coding, math, and multilingual performance.

Grok update with stronger reasoning and a 128K context window.

Databricks' 132B-total / 36B-active open MoE model for code, math, RAG, and enterprise self-hosted workloads.

2.7B-parameter Phi model showing strong reasoning and language understanding at small scale.

Baidu's fourth-generation ERNIE flagship, announced with stronger understanding, generation, reasoning, and memory.

Google's improved multilingual, reasoning, and coding foundation model family introduced at I/O 2023.

The model that brought reliable multi-step reasoning to the mainstream; size never disclosed.