{"meta":{"project":"WhichAI","site":"https://whichai.wiki","apiVersion":"1.0.0","appVersion":"v0.34","license":"CC BY 4.0 (attribution: WhichAI, whichai.wiki)","dataUpdated":"September 8, 2026","benchmarkUpdated":"September 8, 2026","scaleNote":"Artificial Analysis Intelligence Index, 2026 rebased scale (top models ~ 55-66). Snapshot: September 4, 2026 (BenchLM mirror), retrieved September 8, 2026. Specs and prices come from the OpenRouter public model list on the same date. Models nobody has measured carry no index score at all and never enter a ranking: where an estimate exists it is shown separately and labelled unmeasured.","specNote":"Context, price, modality and release dates come from the public OpenRouter model list (retrieved September 8, 2026) and vendor pages. Missing values are not published or not yet verified - shown as -."},"total":118,"count":118,"offset":0,"limit":200,"models":[{"id":"fable-5-1","name":"Claude Fable 5.1","vendor":"Anthropic","family":"claude","status":"public","spec":{"released":"2026-09-01","ctx":"1M","modal":"Text + image + files","priceIn":10,"priceOut":50,"note":"$0.25 per 1M cached input, 75% cheaper than 5.0 · 128K max output"},"access":"Released September 1, 2026 · Claude Pro and Max · API $10/$50 per 1M tokens ($0.25 cached)","tags":["paid","api","prompt-target"],"labels":["writing","reasoning","coding","agents","vision"],"score":{"aa":65.7,"est":false,"cat":{"coding":92,"reasoning":94,"writing":97,"agents":92}},"review":"The highest measured score on this table. Anthropic's September refresh keeps Fable's purpose-built creative-writing strength and adds the thing that actually changes budgets: cache reads at a quarter of the old price, which makes long-document and repeated-context work roughly four times cheaper to run."},{"id":"claude-opus-5","name":"Claude Opus 5","vendor":"Anthropic","family":"claude","status":"public","access":"Released July 24, 2026 · Claude Pro / Max plans · API $5/$25 per 1M tokens (Fast mode $10/$50)","tags":["paid","api","prompt-target"],"labels":["reasoning","coding","agents","long-context"],"spec":{"released":"2026-07-24","ctx":"1M","modal":"Text + vision","priceIn":5,"priceOut":25,"note":"cache hits $0.50/M; Batch -50%; 128K output"},"score":{"aa":63,"est":false,"cat":{"coding":93,"reasoning":94,"writing":88,"agents":95}},"review":"Anthropic's July flagship: 1M context, thinking on by default, computer use, and the Agentic Index lead, at unchanged Opus pricing. The default choice for long agentic runs where you want one model to hold the whole task."},{"id":"fable-5","name":"Claude Fable 5","vendor":"Anthropic","family":"claude","status":"public","access":"Claude Pro / Max plans (usage credits since July 7) · API $10/$50 per 1M tokens","spec":{"ctx":"1M","modal":"Text + vision","priceIn":10,"priceOut":50,"note":"up to 128K output tokens"},"tags":["paid","api","prompt-target"],"labels":["writing","reasoning","agents"],"score":{"aa":62.1,"est":false,"cat":{"coding":89,"reasoning":92,"writing":93,"agents":90}},"review":"Anthropic's newest flagship (Mythos-class) - tops the AA index at max effort and leads creative-writing boards for voice, subtext and character work; briefly export-controlled, now globally available."},{"id":"muse-spark-1-3","name":"Muse Spark 1.3","vendor":"Meta","family":"meta","status":"public","spec":{"released":"2026-09-02","ctx":"1.05M","modal":"Text + image + files + audio + video","priceIn":1.25,"priceOut":4.25,"note":"$0.15 per 1M cached input · up to 940K output tokens · a Contributor tier runs at $0.10/$0.20"},"access":"Released September 2, 2026 · API $1.25/$4.25 per 1M tokens · Contributor tier $0.10/$0.20","tags":["paid","api","prompt-target"],"labels":["agents","vision","reasoning","long-context","value"],"score":{"aa":62.1,"est":false,"cat":{"coding":86,"reasoning":89,"writing":82,"agents":90}},"review":"Meta's second-place finish, and the best value in the top five by a wide margin: a fifth of the frontier input price, natively multimodal across audio and video, and an output window measured in hundreds of thousands of tokens. Three versions in eight weeks says Meta has stopped experimenting with the Muse line and started shipping it."},{"id":"gpt-6","name":"GPT-6 Astra","vendor":"OpenAI","family":"chatgpt","status":"public","spec":{"released":"2026-09-04","ctx":"1.05M","modal":"Text + image + files","priceIn":10,"priceOut":50,"note":"$1 per 1M cached input · 128K max output · a Pro tier and a half-price batch tier ship alongside"},"access":"Released September 4, 2026 · API $10/$50 per 1M tokens ($1 cached) · ChatGPT Plus and Pro · batch tier at $5/$25","tags":["paid","api","prompt-target"],"labels":["reasoning","agents","coding","vision","long-context"],"score":{"aa":61.2,"est":false,"cat":{"coding":91,"reasoning":93,"writing":84,"agents":93}},"review":"OpenAI's next-numbered flagship, shipped September 4 after months of rumour. It ends OpenAI's run of point releases: 1.05M context, native image and file input, and a price that doubles Sol's. Worth knowing before you switch: it scores below both Anthropic flagships on the index, and Sol remains the better value for agentic work at a fifth of the output price."},{"id":"grok-4-6","name":"Grok 4.6","vendor":"xAI","family":"grok","status":"public","spec":{"released":"2026-08-12","ctx":"500K","priceIn":2,"priceOut":6,"speed":"~58 tok/s (below the ~86 tok/s field median)","note":"$0.50 per 1M cached input · reasoning effort: low / medium / high / xhigh"},"access":"Released August 12, 2026 · API $2/$6 per 1M tokens ($0.50 cached) · in Grok Build and Cursor · SuperGrok subscription","tags":["paid","api","prompt-target"],"labels":["coding","agents","reasoning"],"score":{"aa":60.9,"est":false,"cat":{"coding":90,"reasoning":92,"writing":78,"agents":93}},"review":"xAI's August flagship and its first entry in the AA top three (60.9): 70.8% on CursorBench 3.2, 87.0% on VulcanBench v3, and selectable reasoning effort from low to xhigh. Two honest caveats: the 500K context is half what the Claude, GPT and Kimi flagships offer, and measured throughput (~58 tok/s) sits below the field median."},{"id":"kimi-k3","name":"Kimi K3","vendor":"Moonshot AI","family":"kimi","status":"public","access":"API $3/$15 per 1M tokens ($0.30 cached) · free kimi.com chat · open weights on Hugging Face (July 27, Modified MIT)","tags":["free","paid","api","open-weights","prompt-target"],"labels":["coding","reasoning","agents","vision","long-context"],"spec":{"released":"2026-07-16","ctx":"1M","modal":"Text + vision","priceIn":3,"priceOut":15},"score":{"aa":59.7,"est":false,"cat":{"coding":90,"reasoning":91,"writing":78,"agents":94}},"review":"Moonshot's July flagship: 2.8T-parameter MoE, 1M context, native vision, and the best published BrowseComp score (91.2%). The strongest agentic challenger to the US frontier, and the largest model whose weights you can actually download."},{"id":"glm-5-3","name":"GLM-5.3","vendor":"Z.ai","family":"glm","status":"public","spec":{"released":"2026-08-14","note":"743B post-trained base · weights held back pending a staged safety review · price not published"},"access":"Released August 14, 2026 · staged access · open weights NOT released at launch (unlike every previous GLM) · API price not published","tags":["api","prompt-target"],"labels":["coding","reasoning","agents"],"score":{"aa":59.5,"est":false,"cat":{"coding":91,"reasoning":88,"writing":72,"agents":88}},"review":"Z.ai's 743B flagship: AA 59.5, the highest a Chinese lab has scored, with 84.5% on CyberGym and 54.4% on ExploitBench. The break with tradition matters more than the score: weights are held back behind a staged safety review, so for now GLM-5.3 is not the open model its version number suggests. The Flash sibling is."},{"id":"gpt-5-6-sol","name":"GPT-5.6 Sol","vendor":"OpenAI","family":"chatgpt","status":"public","access":"Released July 9, 2026 (rolling out) · API $5/$30 per 1M tokens · up to 750 tok/s on Cerebras","tags":["paid","api","prompt-target"],"labels":["reasoning","agents","coding"],"spec":{"released":"2026-07-09","ctx":"1M","priceIn":5,"priceOut":30,"speed":"up to 750 tok/s (Cerebras)"},"score":{"aa":58.9,"est":false,"cat":{"coding":91,"reasoning":93,"writing":82,"agents":93}},"review":"OpenAI's July flagship: max reasoning effort, programmatic tool calling and an 'ultra mode' that spins up subagents. Since GPT-6 Astra arrived at double the output price, Sol is the better-value half of OpenAI's line for agentic work."},{"id":"gemini-3-8-flash","name":"Gemini 3.8 Flash","vendor":"Google","family":"gemini","status":"public","spec":{"released":"2026-09-02","ctx":"1.05M","modal":"Text + image + files + audio + video","priceIn":0.75,"priceOut":3.75,"note":"$0.075 per 1M cached input · 64K max output"},"access":"Released September 2, 2026 · API $0.75/$3.75 per 1M tokens ($0.075 cached) · free-tier availability not confirmed","tags":["paid","api","prompt-target"],"labels":["coding","agents","speed","value","vision"],"score":{"aa":58.7,"est":false,"cat":{"coding":86,"reasoning":85,"writing":76,"agents":86}},"review":"Google shipped 3.7 Flash on August 13 and 3.8 Flash three weeks later at the same price, 2.7 points higher, with audio and video input added. The best score-per-dollar on the whole table. WhichAI still defaults auto-run to gemini-3.6-flash because that free tier is confirmed and this one is not."},{"id":"qwen-3-8-max-preview","name":"Qwen3.8 Max (preview)","vendor":"Alibaba","family":"qwen","status":"preview","access":"Early preview via Qwen chat and Alibaba Cloud (July 19) · specs not final","tags":["preview","api","prompt-target"],"labels":["reasoning","multilingual"],"spec":{"released":"2026-07-19"},"score":{"aa":58.1,"est":false,"cat":{"coding":80,"reasoning":82,"writing":74,"agents":78}},"review":"Alibaba's next flagship, previewed in July and measured on the August index at a level that puts it among the strongest models with published weights pending. Still labelled preview by Alibaba: treat availability and pricing as provisional even though the score is now a real measurement."},{"id":"gpt-5-5-pro","name":"GPT-5.5 Pro","vendor":"OpenAI","family":"chatgpt","status":"public","access":"ChatGPT Pro ($200/mo) · API","tags":["paid","api","prompt-target"],"labels":["reasoning"],"score":{"aa":58,"est":true,"cat":{"coding":90,"reasoning":91,"writing":82,"agents":87}},"review":"Maximum-reasoning tier of GPT-5.5 - statistically tied with the very top on many boards. Overkill for everyday tasks."},{"id":"glm-5-3-flash","name":"GLM-5.3-Flash","vendor":"Z.ai","family":"glm","status":"public","spec":{"released":"2026-08-26","ctx":"1M","modal":"Text + vision","priceIn":0.15,"priceOut":0.5,"note":"320B-A18B MoE, MIT licence"},"access":"Released August 26, 2026 · open weights under MIT · API $0.15/$0.50 per 1M tokens · free chat at chat.z.ai","tags":["free","paid","api","open-weights","prompt-target"],"labels":["coding","agents","value","vision","long-context"],"score":{"aa":57.5,"est":false,"cat":{"coding":89,"reasoning":84,"writing":70,"agents":86}},"review":"The most capable freely downloadable model on this table: 320B-A18B, MIT licence, 1M multimodal context, and $0.15/$0.50 hosted - roughly a thirtieth of what the frontier charges for a score within six points of it. If one entry explains why the open-weight gap stopped mattering in 2026, it is this one."},{"id":"opus-4-8","name":"Claude Opus 4.8","vendor":"Anthropic","family":"claude","status":"public","spec":{"ctx":"1M","priceIn":5,"priceOut":25},"access":"Claude Pro $20/mo · API","tags":["paid","api","prompt-target"],"labels":["coding","agents","reasoning"],"score":{"aa":57.3,"est":false,"cat":{"coding":92,"reasoning":88,"writing":84,"agents":91}},"review":"The leader on real-repo coding when it shipped (SWE-bench Pro 69.2%) and still a serious software workhorse, now at a price its successors undercut. Worth choosing deliberately rather than by default."},{"id":"muse-spark-1-2","name":"Muse Spark 1.2","vendor":"Meta","family":"meta","status":"public","spec":{"released":"2026-08-06"},"access":"API via Meta Superintelligence Labs (August 6, 2026 refresh of the 1.1 line)","tags":["paid","api","prompt-target"],"labels":["agents","vision","reasoning"],"score":{"aa":56.8,"est":false,"cat":{"coding":80,"reasoning":82,"writing":70,"agents":85}},"review":"Meta's August refresh puts the Muse line back in the top ten (56.8, up 3.6 from 1.1): the strongest showing yet for Meta's proprietary track, and a reminder that the Llama brand is no longer where Meta's frontier work happens."},{"id":"gpt-5-6-terra","name":"GPT-5.6 Terra","vendor":"OpenAI","family":"chatgpt","status":"public","spec":{"released":"2026-07-09","priceIn":2.5,"priceOut":15},"access":"Released July 9, 2026 · API $2.50/$15 per 1M tokens","tags":["paid","api","prompt-target"],"labels":["reasoning","value"],"score":{"aa":56.6,"est":false,"cat":{"coding":86,"reasoning":88,"writing":80,"agents":86}},"review":"The balanced tier of the GPT-5.6 family - GPT-5.5-level quality at half the price. On track to become the everyday ChatGPT default."},{"id":"gpt-5-5","name":"GPT-5.5","vendor":"OpenAI","family":"chatgpt","status":"public","spec":{"released":"2026-04"},"access":"Free (limited messages) · Plus $20/mo · Go ~$8/mo · API","tags":["free","paid","api","prompt-target"],"labels":["coding","reasoning","writing"],"score":{"aa":56.3,"est":false,"cat":{"coding":88,"reasoning":86,"writing":82,"agents":84}},"review":"The April 2026 flagship, still the strongest broad generalist most people can use free: brainstorming, drafting, frontier coding (SWE-bench Verified 88.7%)."},{"id":"gemini-3-7-flash","name":"Gemini 3.7 Flash","vendor":"Google","family":"gemini","status":"public","spec":{"released":"2026-08-13","ctx":"1M","priceIn":0.75,"priceOut":3.75,"note":"introductory price to Dec 31, 2026 - then $1.50/$7.50 · 64K max output"},"access":"Released August 13, 2026 · API $0.75/$3.75 per 1M tokens (introductory, doubles Jan 1, 2027) · model id gemini-3.7-flash · free-tier availability not confirmed","tags":["paid","api","prompt-target"],"labels":["coding","agents","speed","value"],"score":{"aa":56,"est":false,"cat":{"coding":84,"reasoning":82,"writing":74,"agents":84}},"review":"Google's August workhorse, aimed squarely at coding and agents: 85.8% on Terminal-bench 2.1, 97.0% on long-context recall at 128K, and half the price of 3.6 Flash until the end of the year. WhichAI still defaults auto-run to gemini-3.6-flash because that model's free tier is confirmed and this one's is not - switch it in Settings if Google has enabled yours."},{"id":"grok-4-5","name":"Grok 4.5","vendor":"xAI","family":"grok","status":"public","spec":{"released":"2026-07-08","priceIn":2,"priceOut":6,"speed":"~80 tok/s, high token efficiency"},"access":"Released July 8, 2026 · API $2/$6 per 1M tokens · in Grok Build & Cursor · not yet in the EU","tags":["paid","api","prompt-target"],"labels":["coding","agents","value"],"score":{"aa":55.8,"est":false,"cat":{"coding":89,"reasoning":84,"writing":74,"agents":90}},"review":"xAI's first model built specifically for coding and agents: a 1.5T-parameter base trained on real Cursor sessions, at a fraction of flagship prices. Note before you commit: EU availability is still pending."},{"id":"sonnet-5","name":"Claude Sonnet 5","vendor":"Anthropic","family":"claude","status":"public","spec":{"ctx":"1M","priceIn":3,"priceOut":15,"note":"the $2/$10 introductory API price expired Aug 31, 2026"},"access":"FREE tier of claude.ai · API","tags":["free","api","prompt-target"],"labels":["writing","coding","value"],"score":{"aa":55.3,"est":false,"cat":{"coding":84,"reasoning":82,"writing":86,"agents":80}},"review":"The strongest free option on the market - frontier-adjacent writing and coding quality at zero cost. Best zero-budget daily driver for text work."},{"id":"opus-4-7","name":"Claude Opus 4.7","vendor":"Anthropic","family":"claude","status":"legacy","access":"API (being phased down)","tags":["paid","api","legacy","prompt-target"],"labels":["coding","reasoning"],"score":{"aa":55,"est":false,"cat":{"coding":86,"reasoning":84,"writing":80,"agents":84}},"review":"Still top-10 with adaptive reasoning - a capable fallback where 4.8 isn't available."},{"id":"muse-spark-1-1","name":"Muse Spark 1.1","vendor":"Meta","family":"meta","status":"public","spec":{"released":"2026-07-09"},"access":"API via Meta Superintelligence Labs (Meta's first paid model, July 9, 2026)","tags":["paid","api","prompt-target"],"labels":["agents","vision","reasoning"],"score":{"aa":53.2,"est":false,"cat":{"coding":74,"reasoning":76,"writing":66,"agents":80}},"review":"Meta's first proprietary, paid model line (Meta Superintelligence Labs): natively multimodal reasoning with visual chain-of-thought and multi-agent orchestration. A strategic U-turn from open weights."},{"id":"gpt-5-4","name":"GPT-5.4","vendor":"OpenAI","family":"chatgpt","status":"public","access":"ChatGPT Plus · API · inside Perplexity Pro","tags":["paid","api","prompt-target"],"labels":["coding","reasoning"],"score":{"aa":53.1,"est":false,"cat":{"coding":82,"reasoning":80,"writing":76,"agents":78}},"review":"The previous flagship, still very strong and a reliable known quantity - often cheaper per token than its successors."},{"id":"glm-5-2","name":"GLM-5.2","vendor":"Z.ai","family":"glm","status":"public","access":"Open weights · low-cost API hosts","tags":["free","open-weights","api","prompt-target"],"labels":["coding","agents","value"],"score":{"aa":52.6,"est":false,"cat":{"coding":90,"reasoning":80,"writing":68,"agents":84}},"review":"A workhorse open-weight coder: SWE-bench Pro 62.1% and near the top of the React board on Code Arena. Superseded on raw capability by GLM-5.3-Flash, but still the lighter, cheaper option of the two to self-host."},{"id":"gpt-5-6-luna","name":"GPT-5.6 Luna","vendor":"OpenAI","family":"chatgpt","status":"public","spec":{"released":"2026-07-09","priceIn":1,"priceOut":6},"access":"Released July 9, 2026 · API $1/$6 per 1M tokens","tags":["paid","api","prompt-target"],"labels":["speed","value"],"score":{"aa":52.3,"est":false,"cat":{"coding":78,"reasoning":80,"writing":74,"agents":78}},"review":"The fast, affordable member of the 5.6 series - near-flagship scores at commodity prices; strong default for high-volume API work."},{"id":"gemini-3-5-flash","name":"Gemini 3.5 Flash","vendor":"Google","family":"gemini","status":"public","spec":{"ctx":"1M","priceIn":1.5,"priceOut":9},"access":"FREE (most generous free tier) · free API via Google AI Studio","tags":["free","api","prompt-target","auto-run"],"labels":["speed","value","vision"],"score":{"aa":52,"est":false,"cat":{"coding":80,"reasoning":78,"writing":74,"agents":76}},"review":"Remarkable: the newest Flash now outscores the older 3.1 Pro on the AA index - frontier-adjacent quality, free, and the model WhichAI auto-runs with your AI Studio key."},{"id":"gemini-3-6-flash","name":"Gemini 3.6 Flash","vendor":"Google","family":"gemini","status":"public","access":"Free tier in AI Studio and the Gemini app (rate-limited) · API $1.50/$7.50 per 1M tokens","tags":["free","paid","api","prompt-target","auto-run"],"labels":["speed","value"],"spec":{"released":"2026-07-21","ctx":"1M","priceIn":1.5,"priceOut":7.5},"score":{"aa":51.6,"est":false,"cat":{"coding":76,"reasoning":78,"writing":72,"agents":76}},"review":"Google's July refresh of the free workhorse: 3.5 Flash quality with cheaper output ($7.50 vs $9), 1M context, day one in AI Studio, the API and the app. The new default of the Gemini line."},{"id":"gemini-3-1-pro","name":"Gemini 3.1 Pro","vendor":"Google","family":"gemini","status":"public","spec":{"ctx":"1M","priceIn":2,"priceOut":12,"note":"$4/$18 above 200K input tokens"},"access":"Google AI Pro $19.99/mo · API · inside Perplexity Pro","tags":["paid","api","prompt-target"],"labels":["coding","long-context","vision"],"score":{"aa":47.7,"est":false,"cat":{"coding":84,"reasoning":78,"writing":74,"agents":78}},"review":"Led LMArena's coding arena at launch and still the Gemini tier with the deepest Google Workspace integration. On raw capability the Flash line has since overtaken it, so pick this one for the ecosystem, not the score."},{"id":"qwen-3-7-max","name":"Qwen3.7 Max","vendor":"Alibaba","family":"qwen","status":"public","access":"Free chat app · API (closed weights for Max tier)","tags":["free","paid","api","prompt-target"],"labels":["multilingual","coding","reasoning"],"score":{"aa":46.7,"est":false,"cat":{"coding":82,"reasoning":78,"writing":70,"agents":76}},"review":"Alibaba's strongest current model and the top Chinese closed-weights entry on the AA index - outstanding multilingual coverage."},{"id":"kimi-k2-6","name":"Kimi K2.6","vendor":"Moonshot AI","family":"kimi","status":"public","access":"Open weights · API hosts · free chat app","tags":["free","open-weights","api","prompt-target"],"labels":["coding","agents","long-context"],"score":{"aa":45.1,"est":false,"cat":{"coding":88,"reasoning":76,"writing":66,"agents":82}},"review":"Strong open coding (SWE-bench Verified 80.2%) and long agentic tasks - now the smaller sibling of the flagship K3."},{"id":"minimax-m3","name":"MiniMax M3","vendor":"MiniMax","family":"llama","status":"public","access":"Open weights · API hosts","tags":["free","open-weights","api","prompt-target"],"labels":["coding","long-context"],"score":{"aa":44.4,"est":false,"cat":{"coding":82,"reasoning":74,"writing":64,"agents":76}},"review":"MiniMax's flagship, with a reputation for handling very large contexts without falling apart. Comfortably outscored by the current open-weight leaders, so it earns a place on long-document work rather than on general capability."},{"id":"gpt-5-3-codex","name":"GPT-5.3-Codex","vendor":"OpenAI","family":"chatgpt","status":"public","access":"API · Codex tooling","tags":["paid","api","prompt-target"],"labels":["coding","agents"],"score":{"aa":44.3,"est":false,"cat":{"coding":85,"reasoning":72,"writing":55,"agents":84}},"review":"Coding specialist built for agentic dev workflows rather than chat - punches far above its overall index score inside a code editor."},{"id":"deepseek-v4-pro","name":"DeepSeek V4 Pro","vendor":"DeepSeek","family":"deepseek","status":"public","access":"Free chat app · open weights · very cheap API · 1M context default","tags":["free","open-weights","api","prompt-target"],"labels":["reasoning","coding","value"],"score":{"aa":44.3,"est":false,"cat":{"coding":87,"reasoning":82,"writing":68,"agents":76}},"review":"The open reasoning champion (1.6T sparse MoE, 49B active): tops open coding boards at famously low prices - the most popular frontier-adjacent free model worldwide."},{"id":"opus-4-6","name":"Claude Opus 4.6","vendor":"Anthropic","family":"claude","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":["coding"],"score":{"aa":43.7,"est":false,"cat":{"coding":76,"reasoning":74,"writing":72,"agents":72}},"review":"Early-2026 Opus, superseded twice since - retained mainly for pinned enterprise workflows."},{"id":"inkling","name":"Inkling","vendor":"Thinking Machines Lab","family":"llama","status":"public","access":"Open weights (Apache 2.0) on Hugging Face · Tinker API (256K ctx)","tags":["free","open-weights","api","prompt-target"],"labels":["vision","local","long-context"],"spec":{"released":"2026-07-15","ctx":"1M","modal":"Text, image, audio"},"score":{"aa":42.3,"est":false,"cat":{"coding":62,"reasoning":68,"writing":60,"agents":64}},"review":"Thinking Machines' first open release: 975B-parameter MoE (41B active) reasoning natively over text, images and audio, with controllable thinking effort - the new leading US open-weights base for customization."},{"id":"gpt-5-2","name":"GPT-5.2","vendor":"OpenAI","family":"chatgpt","status":"legacy","access":"Still serves some free-tier traffic","tags":["free","legacy","prompt-target"],"labels":["value"],"score":{"aa":42.2,"est":false,"cat":{"coding":68,"reasoning":66,"writing":66,"agents":62}},"review":"Early-2026 default, now superseded - fine for light tasks, noticeably weaker on hard reasoning."},{"id":"mimo-v2-5-pro","name":"MiMo-V2.5-Pro","vendor":"Xiaomi","family":null,"status":"public","access":"Xiaomi cloud & devices (closed)","tags":["paid","info-only"],"labels":["vision","speed"],"score":{"aa":42.2,"est":false,"cat":{"coding":70,"reasoning":70,"writing":60,"agents":66}},"review":"Xiaomi's surprisingly strong in-house line (an Omni sibling handles voice/vision on-device) - a sign of how deep the Chinese frontier bench now runs."},{"id":"kimi-k2-7-code","name":"Kimi K2.7 Code","vendor":"Moonshot AI","family":"kimi","status":"preview","access":"Early access via Moonshot · open weights promised","tags":["preview","open-weights","info-only"],"labels":["coding","agents"],"score":{"aa":42,"est":false,"cat":{"coding":86,"reasoning":70,"writing":55,"agents":82}},"review":"Moonshot's agentic-coding bet - early numbers put it just behind K2.6 overall but ahead on real-repo work."},{"id":"doubao-seed-2-1-pro","name":"Doubao Seed 2.1 Pro","vendor":"ByteDance","family":null,"status":"public","access":"China-focused: served via ByteDance's Volcano Engine cloud only","tags":["paid","api","info-only"],"labels":["enterprise","multilingual"],"score":{"aa":42,"est":true,"cat":{"coding":72,"reasoning":70,"writing":64,"agents":68}},"review":"Built for raw scale - the Doubao family reportedly handles over a hundred trillion token calls a day, mostly inside China's consumer apps. Hard to access from the West."},{"id":"hy3","name":"Hunyuan 3.0 (Hy3)","vendor":"Tencent","family":"llama","status":"public","access":"Open weights (Apache 2.0, released July 6, 2026) · 295B MoE, 21B active · 256K context","tags":["free","open-weights","prompt-target"],"labels":["multilingual","long-context"],"score":{"aa":41.2,"est":false,"cat":{"coding":62,"reasoning":60,"writing":58,"agents":56}},"review":"Tencent's move into the open-weight camp - one of the last big closed Chinese labs to open up, days ago. Solid mid-frontier scores and a permissive license."},{"id":"opus-4-5","name":"Claude Opus 4.5","vendor":"Anthropic","family":"claude","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":["coding"],"score":{"aa":40.8,"est":false,"cat":{"coding":72,"reasoning":70,"writing":70,"agents":68}},"review":"The late-2025 flagship that made agentic coding mainstream; now two generations back."},{"id":"deepseek-v4-flash","name":"DeepSeek V4 Flash","vendor":"DeepSeek","family":"deepseek","status":"public","access":"Free chat app · open weights · cheapest DeepSeek API","tags":["free","open-weights","api","prompt-target"],"labels":["speed","value"],"score":{"aa":40.3,"est":false,"cat":{"coding":74,"reasoning":70,"writing":62,"agents":66}},"review":"The efficient V4 (284B total, 13B active) - most of Pro's quality at a fraction of the compute."},{"id":"glm-5-1","name":"GLM-5.1","vendor":"Z.ai","family":"glm","status":"public","access":"Open weights","tags":["free","open-weights","prompt-target"],"labels":["coding"],"score":{"aa":40.2,"est":false,"cat":{"coding":78,"reasoning":68,"writing":60,"agents":72}},"review":"The previous Z.ai release - still strong, quickly superseded by 5.2."},{"id":"gpt-5-4-mini","name":"GPT-5.4 mini","vendor":"OpenAI","family":"chatgpt","status":"public","access":"API (cheap tier) · serves ChatGPT free-tier overflow","tags":["free","paid","api","prompt-target"],"labels":["speed","value"],"score":{"aa":40,"est":false,"cat":{"coding":66,"reasoning":64,"writing":62,"agents":62}},"review":"Small, quick, cheap - the sensible default for classification, extraction and light drafting at volume."},{"id":"grok-build-0-1","name":"Grok Build 0.1","vendor":"xAI","family":"grok","status":"public","access":"Inside Grok Build (xAI's coding agent/CLI)","tags":["paid","prompt-target"],"labels":["agents","coding"],"score":{"aa":39.8,"est":false,"cat":{"coding":76,"reasoning":62,"writing":45,"agents":80}},"review":"The agent-tuned model behind xAI's Grok Build environment - sessions, skills, plugins and repo-scale work rather than chat."},{"id":"qwen-3-6-plus","name":"Qwen3.6 Plus","vendor":"Alibaba","family":"qwen","status":"public","access":"Free chat app · open-weight siblings · API","tags":["free","open-weights","api","prompt-target"],"labels":["multilingual","coding"],"score":{"aa":39.6,"est":false,"cat":{"coding":72,"reasoning":66,"writing":62,"agents":64}},"review":"Among the strongest open lines for coding and clearly the best multilingual open family."},{"id":"gemini-3-pro","name":"Gemini 3 Pro","vendor":"Google","family":"gemini","status":"public","access":"Google AI Pro · API","tags":["paid","api","prompt-target"],"labels":["vision"],"score":{"aa":39.5,"est":false,"cat":{"coding":70,"reasoning":66,"writing":66,"agents":66}},"review":"Previous Pro generation, still widely deployed across Google products."},{"id":"glm-5","name":"GLM-5","vendor":"Z.ai","family":"glm","status":"public","access":"Open weights","tags":["free","open-weights","prompt-target"],"labels":["coding"],"score":{"aa":39.5,"est":false,"cat":{"coding":76,"reasoning":66,"writing":58,"agents":70}},"review":"The early-2026 base of the GLM-5 family that put Z.ai on the frontier map."},{"id":"qwen-3-7-plus","name":"Qwen3.7 Plus","vendor":"Alibaba","family":"qwen","status":"public","access":"Free chat app · cheap API","tags":["free","api","prompt-target"],"labels":["multilingual","value"],"score":{"aa":39,"est":false,"cat":{"coding":70,"reasoning":66,"writing":62,"agents":64}},"review":"The balanced tier of the 3.7 line - strong value for multilingual work at volume."},{"id":"gpt-5-4-nano","name":"GPT-5.4 nano","vendor":"OpenAI","family":"chatgpt","status":"public","access":"API (cheapest OpenAI tier)","tags":["paid","api"],"labels":["speed","value"],"score":{"aa":38.2,"est":false,"cat":{"coding":60,"reasoning":58,"writing":56,"agents":55}},"review":"OpenAI's smallest current model - surprisingly capable for its price class, built for massive-scale simple tasks."},{"id":"minimax-m2-7","name":"MiniMax M2.7","vendor":"MiniMax","family":"llama","status":"public","access":"Open weights · API hosts","tags":["free","open-weights","api","prompt-target"],"labels":["coding"],"score":{"aa":38.1,"est":false,"cat":{"coding":72,"reasoning":64,"writing":58,"agents":66}},"review":"Strong open coding contender in head-to-head arenas - less known in the West, worth watching."},{"id":"ernie-5-1","name":"ERNIE 5.1","vendor":"Baidu","family":null,"status":"public","access":"Baidu cloud & apps (closed weights)","tags":["paid","api","info-only"],"labels":["multilingual","enterprise"],"score":{"aa":38,"est":true,"cat":{"coding":64,"reasoning":64,"writing":60,"agents":58}},"review":"Baidu's flagship stays closed and China-centric - competitive domestically, rarely benchmarked on Western boards."},{"id":"nemotron-3-ultra","name":"Nemotron 3 Ultra","vendor":"NVIDIA","family":"nemotron","status":"public","spec":{"ctx":"1M","speed":"400+ tok/s on some hosts"},"access":"FREE on OpenRouter (nvidia/nemotron-3-ultra-550b-a55b:free - WhichAI default runner) · open weights · Perplexity Pro","tags":["free","open-weights","api","prompt-target","auto-run"],"labels":["speed","value","reasoning"],"score":{"aa":37.8,"est":false,"cat":{"coding":68,"reasoning":68,"writing":58,"agents":62}},"review":"Strongest US open model and extremely fast (400+ tokens/second on some hosts) - and genuinely free via OpenRouter."},{"id":"grok-4-3","name":"Grok 4.3","vendor":"xAI","family":"grok","status":"public","access":"grok.com paid tiers · API","tags":["paid","api","prompt-target"],"labels":["reasoning"],"score":{"aa":37.6,"est":false,"cat":{"coding":66,"reasoning":68,"writing":60,"agents":64}},"review":"Spring-2026 generation - a solid step over Grok 4.1 on reasoning, now overshadowed by 4.5."},{"id":"qwen-3-6-27b","name":"Qwen3.6-27B","vendor":"Alibaba","family":"qwen","status":"public","access":"Open weights (Apache 2.0) · runs on a good desktop","tags":["free","open-weights","prompt-target"],"labels":["local","multilingual","value"],"score":{"aa":37,"est":false,"cat":{"coding":66,"reasoning":62,"writing":58,"agents":58}},"review":"Possibly the best quality-per-GB you can download today - near mid-frontier scores from a 27B you can run yourself."},{"id":"gpt-5-1","name":"GPT-5.1","vendor":"OpenAI","family":"chatgpt","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":[],"score":{"aa":36.9,"est":false,"cat":{"coding":62,"reasoning":60,"writing":62,"agents":55}},"review":"Late-2025 generation, kept for compatibility - two clear steps behind the current line."},{"id":"gemini-3-5-flash-lite","name":"Gemini 3.5 Flash-Lite","vendor":"Google","family":"gemini","status":"public","access":"API $0.30/$2.50 per 1M tokens","tags":["paid","api","prompt-target"],"labels":["speed","value"],"spec":{"released":"2026-07-21","priceIn":0.3,"priceOut":2.5},"score":{"aa":36.5,"est":false,"cat":{"coding":58,"reasoning":60,"writing":56,"agents":55}},"review":"The routing tier of the July refresh: classification, extraction, tagging and short summaries at commodity prices. Not built for hard reasoning."},{"id":"grok-4-1","name":"Grok 4.1","vendor":"xAI","family":"grok","status":"public","access":"Free tier on X / grok.com · paid from ~$16/mo · API","tags":["free","paid","api","prompt-target"],"labels":["research","long-context"],"score":{"aa":36,"est":true,"cat":{"coding":62,"reasoning":66,"writing":62,"agents":62}},"review":"Direct, up-to-the-minute answers with X data, a Think mode and a 2M-token context; among the lowest measured hallucination rates (~4%)."},{"id":"sonnet-4-6","name":"Claude Sonnet 4.6","vendor":"Anthropic","family":"claude","status":"public","access":"Claude Pro · API","tags":["paid","api","prompt-target"],"labels":["writing","value"],"score":{"aa":35.9,"est":false,"cat":{"coding":68,"reasoning":62,"writing":72,"agents":62}},"review":"The previous balanced default of the Claude line - solid quality at higher limits and lower cost."},{"id":"kimi-k2-5","name":"Kimi K2.5","vendor":"Moonshot AI","family":"kimi","status":"public","access":"Open weights · API hosts","tags":["free","open-weights","api","prompt-target"],"labels":["long-context"],"score":{"aa":35.4,"est":false,"cat":{"coding":72,"reasoning":66,"writing":60,"agents":68}},"review":"The previous Kimi - still a solid open all-rounder with excellent long-context behavior."},{"id":"gpt-5","name":"GPT-5","vendor":"OpenAI","family":"chatgpt","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":[],"score":{"aa":34.7,"est":false,"cat":{"coding":60,"reasoning":58,"writing":60,"agents":52}},"review":"The model that unified the GPT-4/o-series era in August 2025. Historically important; superseded several times since."},{"id":"claude-4-1-opus","name":"Claude Opus 4.1","vendor":"Anthropic","family":"claude","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":[],"score":{"aa":33.7,"est":false,"cat":{"coding":62,"reasoning":60,"writing":64,"agents":55}},"review":"Mid-2025 generation - of historical interest and for compatibility with older pipelines."},{"id":"qwen-3-5-397b","name":"Qwen3.5 397B","vendor":"Alibaba","family":"qwen","status":"public","access":"Open weights (Apache 2.0)","tags":["free","open-weights","prompt-target"],"labels":["multilingual","reasoning"],"score":{"aa":33.7,"est":false,"cat":{"coding":64,"reasoning":62,"writing":56,"agents":56}},"review":"The open MoE flagship of the 3.5 wave - big, capable, Apache-licensed."},{"id":"grok-4","name":"Grok 4","vendor":"xAI","family":"grok","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":[],"score":{"aa":33.3,"est":false,"cat":{"coding":58,"reasoning":62,"writing":52,"agents":56}},"review":"Previous xAI generation - superseded by 4.1's larger context and then by the 4.3/4.5 wave."},{"id":"o3-pro","name":"o3-pro","vendor":"OpenAI","family":"chatgpt","status":"legacy","access":"API (legacy reasoning line)","tags":["paid","api","legacy"],"labels":["reasoning"],"score":{"aa":32.5,"est":false,"cat":{"coding":58,"reasoning":66,"writing":45,"agents":50}},"review":"The 2025 deep-reasoning specialist - slow and dated now, but a milestone in chain-of-thought models."},{"id":"o3","name":"o3","vendor":"OpenAI","family":"chatgpt","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":["reasoning"],"score":{"aa":30.4,"est":false,"cat":{"coding":55,"reasoning":62,"writing":42,"agents":48}},"review":"Once the reasoning benchmark-setter; today mainly of historical and compatibility interest."},{"id":"step-3-7-flash","name":"Step 3.7 Flash","vendor":"StepFun","family":"llama","status":"public","access":"Open weights · API","tags":["free","open-weights","api","prompt-target"],"labels":["speed","multilingual"],"score":{"aa":30.3,"est":false,"cat":{"coding":54,"reasoning":52,"writing":50,"agents":46}},"review":"StepFun's fast open tier - part of the broad Chinese open-weight wave now taking ~45% of OpenRouter traffic."},{"id":"qwen3-coder","name":"Qwen3 Coder","vendor":"Alibaba","family":"qwen","status":"public","spec":{"ctx":"1M"},"access":"Open weights · FREE on OpenRouter (qwen/qwen3-coder:free, 1M context) - WhichAI auto-runs it with a free OpenRouter key","tags":["free","open-weights","api","prompt-target","auto-run"],"labels":["coding","long-context","value"],"score":{"aa":30,"est":true,"cat":{"coding":68,"reasoning":52,"writing":44,"agents":58}},"review":"The strongest free coding route on OpenRouter right now - 1M context, tool use, zero cost within the daily free limits."},{"id":"mistral-medium-3-5","name":"Mistral Medium 3.5","vendor":"Mistral AI","family":"llama","status":"public","access":"Open weights · Le Chat · API","tags":["free","open-weights","api","prompt-target"],"labels":["multilingual","value"],"score":{"aa":29.9,"est":false,"cat":{"coding":56,"reasoning":54,"writing":54,"agents":48}},"review":"Mistral's current best on the index - a newer generation that outscores the bigger, older Large 3. Europe's leading open lab."},{"id":"gemma-4-31b","name":"Gemma 4 31B","vendor":"Google","family":"llama","status":"public","access":"Open weights · runs on consumer hardware","tags":["free","open-weights","prompt-target"],"labels":["local","value"],"score":{"aa":29.4,"est":false,"cat":{"coding":54,"reasoning":52,"writing":54,"agents":46}},"review":"Google's open family - one of the best models you can run locally on a single GPU."},{"id":"gemini-3-flash","name":"Gemini 3 Flash","vendor":"Google","family":"gemini","status":"public","access":"Free tier · cheap API","tags":["free","api","prompt-target"],"labels":["speed","value"],"score":{"aa":27.4,"est":false,"cat":{"coding":52,"reasoning":48,"writing":50,"agents":46}},"review":"The earlier fast tier - inexpensive and quick, now clearly behind 3.5 Flash."},{"id":"haiku-4-5","name":"Claude Haiku 4.5","vendor":"Anthropic","family":"claude","status":"public","access":"All Claude tiers · cheapest Claude via API","tags":["free","paid","api","prompt-target"],"labels":["speed","value"],"score":{"aa":27,"est":true,"cat":{"coding":52,"reasoning":45,"writing":55,"agents":45}},"review":"Fastest and cheapest of the line - quick answers, classification, high-volume light tasks."},{"id":"gemini-2-5-pro","name":"Gemini 2.5 Pro","vendor":"Google","family":"gemini","status":"legacy","access":"API (legacy, paid-only since April 2026)","tags":["paid","api","legacy"],"labels":["long-context"],"score":{"aa":25.8,"est":false,"cat":{"coding":50,"reasoning":48,"writing":48,"agents":44}},"review":"The 2025 flagship, superseded by the Gemini 3 line."},{"id":"nemotron-3-super","name":"Nemotron 3 Super","vendor":"NVIDIA","family":"nemotron","status":"public","access":"FREE on OpenRouter (nvidia/nemotron-3-super-120b-a12b:free) · open weights","tags":["free","open-weights","api","prompt-target","auto-run"],"labels":["speed","value"],"score":{"aa":25.4,"est":false,"cat":{"coding":48,"reasoning":46,"writing":42,"agents":40}},"review":"The 120B sibling - lighter, faster, still free; good for volume tasks."},{"id":"gemini-3-1-flash-lite","name":"Gemini 3.1 Flash-Lite","vendor":"Google","family":"gemini","status":"public","access":"Free API (highest free rate limits: 15 RPM, 1,000/day)","tags":["free","api","prompt-target"],"labels":["speed","value"],"score":{"aa":25,"est":false,"cat":{"coding":48,"reasoning":44,"writing":46,"agents":42}},"review":"The volume workhorse: the most generous free API quota of any capable model."},{"id":"deepseek-v3-2","name":"DeepSeek V3.2","vendor":"DeepSeek","family":"deepseek","status":"legacy","access":"Open weights · cheap API","tags":["free","open-weights","api","legacy"],"labels":["value"],"score":{"aa":24.7,"est":false,"cat":{"coding":50,"reasoning":46,"writing":42,"agents":40}},"review":"The 2025 generation that made ultra-cheap frontier-adjacent AI mainstream; superseded by V4."},{"id":"gpt-oss-120b","name":"gpt-oss-120b","vendor":"OpenAI","family":"llama","status":"public","access":"Open weights · free/cheap API hosts","tags":["free","open-weights","api","prompt-target"],"labels":["local","value"],"score":{"aa":23.8,"est":false,"cat":{"coding":45,"reasoning":44,"writing":40,"agents":38}},"review":"OpenAI's open-weight release - modest by 2026 standards but fully downloadable and permissively licensed."},{"id":"command-a-plus","name":"Command A+","vendor":"Cohere","family":null,"status":"public","access":"Enterprise API · open weights for research · private deployments","tags":["paid","api","open-weights","info-only"],"labels":["enterprise","multilingual"],"score":{"aa":22.5,"est":false,"cat":{"coding":42,"reasoning":40,"writing":44,"agents":40}},"review":"Enterprise-focused: retrieval-augmented answers, data residency, private deployment - chosen for compliance more than raw benchmarks."},{"id":"k-exaone","name":"K-Exaone","vendor":"LG AI Research","family":null,"status":"public","access":"Korean cloud ecosystem · some open releases","tags":["paid","info-only"],"labels":["multilingual","enterprise"],"score":{"aa":22.1,"est":false,"cat":{"coding":46,"reasoning":44,"writing":42,"agents":38}},"review":"Korea's leading domestic line - strong on Korean-language benchmarks, part of the sovereign-AI wave."},{"id":"nova-2","name":"Nova 2","vendor":"Amazon","family":null,"status":"public","access":"AWS Bedrock (enterprise)","tags":["paid","api","info-only"],"labels":["enterprise","value"],"score":{"aa":22,"est":true,"cat":{"coding":44,"reasoning":42,"writing":42,"agents":42}},"review":"Amazon's in-house line, priced aggressively inside Bedrock - the default for AWS-native companies, not a benchmark leader."},{"id":"gemma-4-12b","name":"Gemma 4 12B","vendor":"Google","family":"llama","status":"public","access":"Open weights · laptop-friendly","tags":["free","open-weights","prompt-target"],"labels":["local","speed"],"score":{"aa":21.8,"est":false,"cat":{"coding":42,"reasoning":40,"writing":44,"agents":34}},"review":"The mid-size Gemma - a favorite for private, offline use on ordinary hardware."},{"id":"grok-code-fast-1","name":"Grok Code Fast 1","vendor":"xAI","family":"grok","status":"public","access":"API (very cheap) · common inside IDE integrations","tags":["paid","api","prompt-target"],"labels":["coding","speed","value"],"score":{"aa":21.6,"est":false,"cat":{"coding":55,"reasoning":35,"writing":25,"agents":48}},"review":"Tiny, fast coding model - autocompletion-class work at near-zero cost, not for hard problems."},{"id":"mercury-2","name":"Mercury 2","vendor":"Inception Labs","family":null,"status":"public","access":"API","tags":["paid","api","info-only"],"labels":["speed"],"score":{"aa":21.4,"est":false,"cat":{"coding":52,"reasoning":44,"writing":42,"agents":40}},"review":"A diffusion language model - generates text in parallel rather than token-by-token, hitting extreme speeds. The most interesting architectural outsider on the board."},{"id":"deepseek-r1","name":"DeepSeek R1","vendor":"DeepSeek","family":"deepseek","status":"legacy","access":"Open weights","tags":["free","open-weights","legacy"],"labels":["reasoning"],"score":{"aa":20.1,"est":false,"cat":{"coding":42,"reasoning":50,"writing":35,"agents":32}},"review":"The January 2025 release that shocked the market by open-sourcing o1-class reasoning. Historic; superseded."},{"id":"mistral-small-4","name":"Mistral Small 4","vendor":"Mistral AI","family":"llama","status":"public","access":"Open weights (Apache 2.0) · laptop-class","tags":["free","open-weights","prompt-target"],"labels":["local","value"],"score":{"aa":19.6,"est":false,"cat":{"coding":40,"reasoning":36,"writing":40,"agents":30}},"review":"Compact European open model - a common pick for private, on-premise deployments."},{"id":"gpt-4-1","name":"GPT-4.1","vendor":"OpenAI","family":"chatgpt","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":["long-context"],"score":{"aa":19.4,"est":false,"cat":{"coding":42,"reasoning":35,"writing":42,"agents":30}},"review":"2025 workhorse with a then-notable 1M context. Cheap, stable, far from frontier."},{"id":"trinity-large","name":"Trinity Large","vendor":"Arcee AI","family":"llama","status":"public","access":"Open weights","tags":["free","open-weights","prompt-target"],"labels":["local","value"],"score":{"aa":18.2,"est":false,"cat":{"coding":46,"reasoning":44,"writing":42,"agents":40}},"review":"A rare US startup training capable open models from scratch - the 'Trinity' line is the credible American answer to the Chinese open wave."},{"id":"mistral-large-3","name":"Mistral Large 3","vendor":"Mistral AI","family":"llama","status":"public","access":"Open weights · Le Chat free tier · API","tags":["free","open-weights","api","prompt-target"],"labels":["multilingual"],"score":{"aa":15.9,"est":false,"cat":{"coding":36,"reasoning":32,"writing":38,"agents":28}},"review":"The 675B MoE that was Europe's flagship in 2025 - the index now places it well behind newer, smaller Mistral releases."},{"id":"gpt-oss-20b","name":"gpt-oss-20b","vendor":"OpenAI","family":"llama","status":"public","access":"Open weights · runs on a laptop","tags":["free","open-weights","prompt-target"],"labels":["local","speed"],"score":{"aa":14.9,"est":false,"cat":{"coding":32,"reasoning":30,"writing":30,"agents":25}},"review":"The small sibling - one of the easiest capable models to run fully offline on consumer hardware."},{"id":"llama-4-maverick","name":"Llama 4 Maverick","vendor":"Meta","family":"llama","status":"public","access":"Open weights · FREE fast inference on Groq","tags":["free","open-weights","api","prompt-target","auto-run"],"labels":["local","speed","value"],"score":{"aa":14.3,"est":false,"cat":{"coding":34,"reasoning":30,"writing":36,"agents":26}},"review":"The larger Llama 4 - near-instant on Groq and fine for drafts, but the index is honest: it sits far below the 2026 open leaders."},{"id":"gemini-2-5-flash","name":"Gemini 2.5 Flash","vendor":"Google","family":"gemini","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":["speed"],"score":{"aa":14.1,"est":false,"cat":{"coding":32,"reasoning":28,"writing":32,"agents":26}},"review":"2025's fast tier, kept for compatibility."},{"id":"ling-2-6-flash","name":"Ling 2.6 Flash","vendor":"Ant Group","family":"llama","status":"public","access":"Open weights","tags":["free","open-weights","prompt-target"],"labels":["speed","value"],"score":{"aa":14.1,"est":false,"cat":{"coding":32,"reasoning":28,"writing":28,"agents":24}},"review":"Ant Group's efficient open line - built for cheap, fast serving rather than peak intelligence."},{"id":"jamba-large","name":"Jamba Large","vendor":"AI21 Labs","family":null,"status":"public","access":"Open weights · API","tags":["free","open-weights","api","info-only"],"labels":["long-context","enterprise"],"score":{"aa":12,"est":true,"cat":{"coding":26,"reasoning":24,"writing":30,"agents":20}},"review":"AI21's hybrid SSM-Transformer line - memory-efficient very-long-context processing for enterprises, a different trade-off than pure Transformers."},{"id":"sarvam-105b","name":"Sarvam 105B","vendor":"Sarvam AI","family":null,"status":"public","access":"Open weights · Indian cloud","tags":["free","open-weights","info-only"],"labels":["multilingual"],"score":{"aa":11.9,"est":false,"cat":{"coding":24,"reasoning":22,"writing":26,"agents":18}},"review":"India's sovereign-AI flagship - optimized for Indic languages, a milestone for AI outside the US-China duopoly."},{"id":"claude-3-opus","name":"Claude 3 Opus","vendor":"Anthropic","family":"claude","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":["writing"],"score":{"aa":11.8,"est":false,"cat":{"coding":25,"reasoning":22,"writing":40,"agents":15}},"review":"The 2024 model that first made Claude famous for prose quality. Retired from frontier duty."},{"id":"gpt-4o","name":"GPT-4o","vendor":"OpenAI","family":"chatgpt","status":"legacy","access":"API (legacy)","tags":["paid","api","legacy"],"labels":["vision"],"score":{"aa":11.2,"est":false,"cat":{"coding":28,"reasoning":22,"writing":35,"agents":18}},"review":"The 2024-25 multimodal workhorse. Historically important, no longer competitive on quality - kept for compatibility."},{"id":"ministral-3-14b","name":"Ministral 3 14B","vendor":"Mistral AI","family":"llama","status":"public","access":"Open weights · edge/on-device","tags":["free","open-weights"],"labels":["local","speed"],"score":{"aa":11.1,"est":false,"cat":{"coding":26,"reasoning":22,"writing":26,"agents":18}},"review":"Mistral's edge line - built to run on phones and small boxes, not to win benchmarks."},{"id":"llama-4-scout","name":"Llama 4 Scout","vendor":"Meta","family":"llama","status":"public","access":"Open weights · very long context","tags":["free","open-weights","prompt-target"],"labels":["local","long-context"],"score":{"aa":10,"est":false,"cat":{"coding":26,"reasoning":22,"writing":30,"agents":20}},"review":"Llama 4's efficient sibling with an enormous context window - useful locally, weak on hard tasks."},{"id":"llama-3-1-405b","name":"Llama 3.1 405B","vendor":"Meta","family":"llama","status":"legacy","access":"Open weights","tags":["free","open-weights","legacy"],"labels":["local"],"score":{"aa":8.5,"est":false,"cat":{"coding":22,"reasoning":20,"writing":26,"agents":16}},"review":"2024's open giant - a milestone for open AI, now outclassed by far smaller models."},{"id":"lfm-2-5","name":"LFM2.5","vendor":"Liquid AI","family":null,"status":"public","access":"Open weights · edge/on-device","tags":["free","open-weights","info-only"],"labels":["local","speed"],"score":{"aa":8.3,"est":false,"cat":{"coding":18,"reasoning":16,"writing":18,"agents":12}},"review":"'Liquid' non-Transformer architecture squeezing real capability into phone-sized models - watch this space for on-device AI."},{"id":"granite-4","name":"Granite 4.0","vendor":"IBM","family":null,"status":"public","access":"Open weights (Apache 2.0) · watsonx","tags":["free","open-weights","info-only"],"labels":["enterprise","local"],"score":{"aa":8,"est":true,"cat":{"coding":22,"reasoning":18,"writing":20,"agents":16}},"review":"IBM's open enterprise family (hybrid Mamba architecture, tiny-to-mid sizes) - built for governed, on-premise business use, not leaderboards."},{"id":"reka-flash-3","name":"Reka Flash 3","vendor":"Reka AI","family":null,"status":"public","access":"Open weights","tags":["free","open-weights","info-only"],"labels":["local","vision"],"score":{"aa":8,"est":true,"cat":{"coding":20,"reasoning":18,"writing":20,"agents":14}},"review":"Small multimodal open model from an ex-DeepMind team - efficient vision-language work on modest hardware."},{"id":"solar-pro-2","name":"Solar Pro 2","vendor":"Upstage","family":null,"status":"public","access":"API · Korean ecosystem","tags":["paid","api","info-only"],"labels":["multilingual","value"],"score":{"aa":7.8,"est":false,"cat":{"coding":20,"reasoning":16,"writing":20,"agents":14}},"review":"Upstage's efficient Korean-market model - document AI and enterprise Korean text are its home turf."},{"id":"gemma-3-27b","name":"Gemma 3 27B","vendor":"Google","family":"llama","status":"legacy","access":"Open weights","tags":["free","open-weights","legacy"],"labels":["local"],"score":{"aa":7.4,"est":false,"cat":{"coding":20,"reasoning":16,"writing":24,"agents":12}},"review":"The 2025 open generation, superseded by Gemma 4."},{"id":"llama-3-3-70b","name":"Llama 3.3 70B","vendor":"Meta","family":"llama","status":"legacy","access":"Open weights · FREE on Groq (WhichAI default runner)","tags":["free","open-weights","api","legacy","prompt-target","auto-run"],"labels":["local","speed"],"score":{"aa":7,"est":true,"cat":{"coding":20,"reasoning":16,"writing":24,"agents":14}},"review":"Older but dependable - the default model WhichAI auto-runs with a free Groq key."},{"id":"falcon-h1","name":"Falcon H1","vendor":"TII (UAE)","family":null,"status":"public","access":"Open weights","tags":["free","open-weights","info-only"],"labels":["local"],"score":{"aa":7,"est":true,"cat":{"coding":18,"reasoning":16,"writing":18,"agents":12}},"review":"The UAE's hybrid-architecture open family - a persistent, well-funded outsider rather than a frontier contender."},{"id":"olmo-2","name":"OLMo 2","vendor":"Ai2","family":null,"status":"public","access":"Fully open: weights, data and training code","tags":["free","open-weights","info-only"],"labels":["local"],"score":{"aa":5,"est":true,"cat":{"coding":14,"reasoning":12,"writing":16,"agents":8}},"review":"The most transparent model line in existence - everything is published, including training data. Built for science, not benchmarks."},{"id":"phi-4","name":"Phi-4","vendor":"Microsoft","family":"llama","status":"public","access":"Open weights · laptop-friendly","tags":["free","open-weights","prompt-target"],"labels":["local","value"],"score":{"aa":4.9,"est":false,"cat":{"coding":15,"reasoning":14,"writing":16,"agents":8}},"review":"Microsoft's small open research line - historically interesting for data-quality tricks, no longer competitive."},{"id":"mythos-5-1","name":"Claude Mythos 5.1","vendor":"Anthropic","family":"claude","status":"public","spec":{"released":"2026-09-01"},"access":"Released September 1, 2026 alongside Fable 5.1 · availability varies by plan","tags":["paid","api","prompt-target"],"labels":["reasoning","agents"],"score":{"aa":null,"guess":62,"est":true,"cat":{"coding":90,"reasoning":93,"writing":86,"agents":92}},"review":"Shipped the same day as Fable 5.1 and aimed at technical and scientific work (52.6% on Terminal-Bench Science at launch). No index score published yet, so the figure here is an unmeasured estimate and this entry never enters a ranking."},{"id":"mythos-5","name":"Claude Mythos 5","vendor":"Anthropic","family":null,"status":"private","access":"Approved organizations only (Project Glasswing); US government-cleared, expanding","tags":["private","info-only"],"labels":["reasoning","enterprise"],"score":{"aa":null,"guess":61,"est":true,"cat":{"coding":90,"reasoning":94,"writing":88,"agents":90}},"review":"Same underlying model as Fable 5 with dual-use safeguards lifted for vetted organizations; deployed on critical infrastructure in 15+ countries."},{"id":"gemini-3-ultra","name":"Gemini 3 Ultra","vendor":"Google","family":null,"status":"rumored","access":"Reported as upcoming - no official availability","tags":["rumored","info-only"],"labels":["reasoning","long-context"],"score":{"aa":null,"guess":56,"est":true,"cat":{"coding":86,"reasoning":88,"writing":80,"agents":85}},"review":"Reported top tier above Gemini 3.1 Pro with wider API access said to be coming; Google has published nothing definitive - treat specs as rumor."},{"id":"gemini-3-5-pro","name":"Gemini 3.5 Pro","vendor":"Google","family":null,"status":"rumored","access":"Not announced. Release-timing reports around the Kimi K3 launch called it imminent; nothing official from Google","tags":["rumored","info-only"],"labels":["reasoning","coding"],"score":{"aa":null,"guess":55,"est":true,"cat":{"coding":88,"reasoning":89,"writing":80,"agents":86}},"review":"The widely expected Pro refresh of the Gemini line. Third-party release trackers treated it as imminent in late July; until Google speaks, everything here (including the name) is rumor and the score is a placeholder estimate."},{"id":"grok-5","name":"Grok 5","vendor":"xAI","family":null,"status":"rumored","access":"Not released - repeatedly delayed (prediction markets price a near-term launch low)","tags":["rumored","info-only"],"labels":["reasoning"],"score":{"aa":null,"guess":58,"est":true,"cat":{"coding":88,"reasoning":90,"writing":78,"agents":88}},"review":"xAI's long-promised next flagship. Timelines have slipped more than once; any spec you read is speculation."},{"id":"llama-5","name":"Llama 5","vendor":"Meta","family":null,"status":"rumored","access":"Not announced - no weights, no model card","tags":["rumored","info-only"],"labels":["local"],"score":{"aa":null,"guess":45,"est":true,"cat":{"coding":74,"reasoning":72,"writing":68,"agents":70}},"review":"Persistent rumors, zero official signal: no Meta announcement and no Hugging Face weights. Any 'Llama 5 review' online is premature or invented. Meta's proprietary Muse line now gets the frontier attention."},{"id":"meta-ai","name":"Meta AI","vendor":"Meta","family":"meta","status":"public","access":"FREE - meta.ai and inside WhatsApp, Instagram, Messenger","tags":["free","prompt-target"],"labels":["value"],"score":{"aa":null,"est":true,"cat":null},"review":"Not a model but an assistant woven into Meta's apps, now increasingly powered by the Muse line with real-time info via search. Ubiquitous and free."},{"id":"copilot","name":"Copilot (GPT-5 family)","vendor":"Microsoft","family":"copilot","status":"public","access":"Free tier · Copilot Pro $20/mo · built into Windows & Microsoft 365","tags":["free","paid","prompt-target"],"labels":["enterprise"],"score":{"aa":null,"est":true,"cat":null},"review":"Not a model but a delivery vehicle: OpenAI GPT-5-family models woven into Word, Excel, PowerPoint, Teams and Windows."},{"id":"sonar","name":"Perplexity Sonar","vendor":"Perplexity AI","family":"perplexity","status":"public","access":"FREE: unlimited basic cited searches + 5 Pro searches/day","tags":["free","api","prompt-target"],"labels":["research","speed"],"score":{"aa":null,"est":true,"cat":null},"review":"Search-answer engine rather than a chat model: best-in-class citation accuracy - a smarter, cited replacement for googling."},{"id":"sonar-pro","name":"Perplexity Sonar Pro + Deep Research","vendor":"Perplexity AI","family":"perplexity","status":"public","access":"Perplexity Pro $20/mo (includes GPT, Claude and Gemini flagships inside)","tags":["paid","api","prompt-target"],"labels":["research","long-context"],"score":{"aa":null,"est":true,"cat":null},"review":"Deeper retrieval, 20 Deep Research runs/day, and frontier models selectable inside one subscription."}]}