Models

Model rankings

See where capabilities stand in the same evaluation, then explore works and uses. Scores do not cover every task; high and max are tested reasoning settings, and Pi is the tool configuration used for coding comparisons. Explore works · Explore uses

Artificial Analysis intelligence comparison

Source read · · v5.0.1

Source leaderboard ↗

Intelligence Index · From October 29, 2026 this board reads Artificial Analysis Intelligence Index v5.0 in full (the methodology page names v5.0 the same day): GDPval-AA exits, FrontierCode 2.0 (132 tasks from real open-source repositories, scored by maintainers' willingness to merge, blocking check failures score zero) weighs 15%, Terminal-Bench 4.0 weighs 10%, and knowledge work / coding / science / general each weigh 25%, with several evaluations on private data. Each model still uses its highest tested setting with fallback conditions retained; v5.0 and v4.3.2 scores cannot be compared across versions. Kimi K3, Qwen3.8-Flash-Next and Nemotron 3.5 Bolt 74B-A9B have no scored row on the v5.0 source table and leave the board (their last v4.3.2 values 44/40/21 remain in records); GLM-5.3-Flash first shows a v5.0 score on the October 31 re-read and joins at 28 (effort not named; its v4.3.2 42 stays in records); GPT-6.1 Terra now appears on the source table and joins under this board's rules; Gemini 4 Carbon (announced October 28) leads at 59 (high) without general availability yet. For Grok 4.7 under v5.0, high is taken; on the November 4 source table xhigh also reads 35, tying high, so high stays. Compare within one evaluator and one version; the index mainly measures English text tasks — judge Chinese and image capabilities separately. A full re-read on October 31: the other 28 rows are unchanged and GLM-5.3-Flash joins at 28. A full re-read on November 2: all 30 rows keep their scores and settings; new uncatalogued source rows (such as DeepSeek V4 Pro 0813 and Inkling) remain outside this board under the standing note. A full re-read on November 3: all 30 rows keep their scores and settings; new uncatalogued rows (Nanbeige4.3-33B, K2 Horizon, the Qwen3.5 series and others) remain outside this board under the standing note; the Grok 4.7 row's organization label now reads SpaceXAI on the source page — a display label only, not recorded here. Prior records stand: Mistral Large 4 weights opened on October 29, the source table retired the Preview row, and the released weights join at high 26 (medium 23), with the preview-era 25 kept in records; other new rows such as Mistral Small 4 are not catalogued. A full re-read on November 4: 25 rows keep their scores and settings while 5 rows gain +1 (Carbon 60, Opus 5.5 59, Muse Wonder 44, DeepSeek 1017 33, Ling 1019 30) — the pre-reread “Carbon and Opus tied for first” becomes “Carbon alone at the top”; new uncatalogued rows (Nanbeige4.3-33B, K2 Horizon, Inkling Small and others) remain outside this board under the standing note. A full re-read on November 5: all 30 rows keep their scores and settings; Hy4 joins the source table at 35 (reasoning version, effort not named in the source table), placed between Grok 4.7's 35 (high) and Qwen3.8 Max's 34; new uncatalogued rows remain outside this board under the standing note. A full re-read on November 6: all 31 rows keep their scores and settings; new uncatalogued rows remain outside this board under the standing note. A full re-read on November 8: all 31 rows keep their scores and settings; new uncatalogued rows (Inkling 2 Small and others) remain outside this board under the standing note. A full re-read on November 9: 30 rows keep their scores and settings; Claude Sonnet 5.5 (max) rises from 54 to 55, tying GPT-6.1 Sol (max), with the source table placing Sol before Sonnet; new uncatalogued rows (Inkling 2 Small, Command A+ and others) remain outside this board under the standing note. A full re-read on November 10: 30 of the 31 catalogued rows keep their scores and settings — the source table retired Claude Fable 5.1 and now lists Claude Fable 5.2 (max 65 at the top, xhigh/high 62, medium 56, low 54) and Muse Ember (max 36, xhigh 34); the Fable 5.1 row is preserved in records (xhigh with fallback 50, its source row retired) following the Spark 1.3 precedent; new uncatalogued rows (Fable 5.2, Muse Ember, Agnes 2.5 Pro Theta and others) remain outside this board under the standing note. Later the same morning: Claude Fable 5.2 was catalogued and joins this board (max 65 at the top, scope 31→32); Muse Ember and the other uncatalogued rows stay outside under the standing note. The same evening: after reading Meta's official developer post, Muse Ember was catalogued and joins this board (max 36, tied with Claude Haiku 5.5; scope 32→33); the official page itself reports 36 on the AA intelligence index (xhigh 34 comes from our November 10 morning full-table read); other uncatalogued rows stay outside under the standing note. A full re-read on November 11: the listed row count checks out at 31 — the earlier “scope 32→33” running note disagrees with the list, a counting drift corrected here (the 31 rows include the Fable 5.1 row preserved under the Spark 1.3 precedent); all 31 catalogued rows keep their highest-tested scores and settings, with Fable 5.2 (max 65) still at the top; the source table adds medium 58 and low 54 rows for Gemini 4 Carbon (our row stays high 60); Carbon became generally available on November 10 — the earlier "without general availability yet" phrase is that day's record, not the current state; new uncatalogued rows remain outside under the standing note. A full re-read on November 13 (direct HTML fetch; web_reader returned a cached copy that predates Hydrogen, so the direct fetch was verified and used): the catalogued rows actually number 33 — the November 11 note saying the listed row count checks out at 31 disagrees with the list, a second counting drift (the Nemotron 3.5 Super VL and Muse Glimmer rows were appended in multi-line format and missed by the line-based count), corrected here. All 33 catalogued rows keep their highest-tested scores and settings. The source table adds four rows for Gemini 4 Hydrogen (xhigh 54 / high 51 / medium 41 / low 27), released November 12 and catalogued the same round, joining at its highest tested setting xhigh 54 (scope 33→34). The methodology page now labels v5.0.1, and its version history states 5.0.1 (November 2026 to current) only "incorporated grading robustness improvements from the CritPt authors and regraded all models on CritPt v2"; every row's displayed score matches the November 11 read, so board continuity is unaffected. New uncatalogued rows (Solar Pro 4, Apodex 1.1, JT-4.1 Flash 236B A21B and others) remain outside under the standing note. A full re-read on November 14 (direct HTML fetch): all 34 catalogued rows keep their highest-tested scores and settings, with Fable 5.2 (max 65) still at the top; rows without a named effort stay as recorded; new uncatalogued rows (Inkling 2 Small now showing 31 on the source table and others) remain outside under the standing note. An evening re-read on November 14 (direct HTML fetch): all 34 catalogued rows keep their highest-tested scores and settings, and the row set matches the morning read; uncatalogued rows such as Inkling 2 Small (31) and Qwen3.8-Flash-Next (29) remain outside under the standing note. A full re-read on November 15 (direct HTML fetch): all 34 catalogued rows keep their highest-tested scores and settings, with Fable 5.2 (max 65) still at the top; the source table now renders one row per model with the effort labeled inline (the same page on November 13 still listed Hydrogen as four per-effort rows) — a display-layout change that does not affect this board's per-row values; uncatalogued rows such as Inkling 2 Small (31), Nanbeige4.3-33B (20), JT-4.1 Flash (19), MiniMax-M3 (19), K2 Horizon (18), Solar Pro 4 (14) and Apodex 1.1 (14) remain outside under the standing note. A full re-read on November 16 (direct HTML fetch): all 34 catalogued rows keep their highest-tested scores and settings, with Fable 5.2 (max 65) still at the top; uncatalogued rows (Inkling 2 Small 31, Nanbeige4.3-33B 20, JT-4.1 Flash 236B A21B 19, MiniMax-M3 19, K2 Horizon 18, Solar Pro 4 14, Apodex 1.1 14) remain outside under the standing note; the source table now shows an unscored Ling 3.1 Flash row (distinct from the catalogued Ling 3.1 Flash 1019), left out under the standing note. An evening re-read on November 16 (direct HTML fetch): all 34 catalogued rows keep their highest-tested scores and settings; the source table now shows a Qwen4 Max row (43, effort not named in the source table) — the model shipped November 16 after a limited qwen4-max-preview from November 11; this round read the official page's title and release facts directly, recorded figures from the IT之家 relay, then catalogued the model and joined it to this board at 43 under the standing rules (scope 34→35); the Grok 4.7 source row now displays Xhigh 35 (previously recorded as high and xhigh tied at 35 — value unchanged, this board's row note stands); new uncatalogued rows (LongCat 2.0 and LongCat Flash Lite without intelligence scores, Inkling Xhigh 12 and others) remain outside under the standing note; the unscored Ling 3.1 Flash row persists. A full re-read on November 17 (direct HTML fetch): all 35 catalogued rows keep their highest-tested scores and settings, with Fable 5.2 (max 65) still at the top and Qwen4 Max unchanged at 43; the source table retired the Qwen3.8 Max (0902) (34) and Gemini 3.8 Flash (high 31) rows (both marked deprecated in the page's embedded model metadata), and both stay on this board under the Fable 5.1 / Spark 1.3 precedent; the unscored Ling 3.1 Flash row seen on November 16 is gone; unscored rows such as Qwen3.8 2.4T A95B and GPT-6 Sol Daybreak Blue (max) remain outside under the standing note; other uncatalogued rows (Inkling 2 Small 31, Nanbeige4.3-33B 20, JT-4.1 Flash 236B A21B 19 and others) remain outside under the standing note.

Preparing filters; the current board is available to read. 35 model configurations · Filters retain the original rank

Scroll sideways for scores and evaluations; model names stay on the left.

Artificial Analysis intelligence comparison · Intelligence Index
RankModel / test configurationCompanyScoreExplore
1Claude Fable 5.2maxAnthropic65Evaluations →
2Gemini 4 CarbonhighGoogle DeepMind60Evaluations →
3Claude Opus 5.5max with fallbackAnthropic59Evaluations →
4GPT-6 AstramaxOpenAI57Evaluations →
5GPT-6.1 SolmaxOpenAI55Evaluations →
5Claude Sonnet 5.5max with fallbackAnthropic55Evaluations →
7Gemini 4 HydrogenxhighGoogle DeepMind54Evaluations →
8GPT-6.1 TerramaxOpenAI51Evaluations →
9Gemini 4 ArgonhighGoogle DeepMind50Evaluations →
9Claude Fable 5.1xhigh with fallbackAnthropic50Evaluations →
9Muse Wonder FullmaxMeta50Evaluations →
12Kimi K3.1 ProReasoning version; effort not named in source tableMoonshot AI / 月之暗面46Evaluations →
13Muse WondermaxMeta44Evaluations →
14Qwen4 MaxReasoning version; effort not named in source tableAlibaba / Qwen43Evaluations →
15Claude Haiku 5.5maxAnthropic36Evaluations →
15Muse EmbermaxMeta36Evaluations →
17Grok 4.7high · tied with xhigh (both 35) under v5.0SpaceXAI35Evaluations →
17Tencent Hy4Reasoning version; effort not named in source tableTencent Hunyuan / 腾讯混元35Evaluations →
19Qwen3.8 Max (0902)0902 snapshot; effort not named in source tableAlibaba / Qwen34Evaluations →
20DeepSeek V4.1 Flash 1017maxDeepSeek33Evaluations →
21GLM-5.3maxZ.ai / 智谱32Evaluations →
22Gemini 3.8 FlashhighGoogle DeepMind31Evaluations →
23MiMo-V2.6-ProEffort not named in source tableXiaomi MiMo / 小米30Evaluations →
23Ling 3.1 Flash 1019Reasoning version; effort not named in source tableInclusionAI30Evaluations →
25MiniMax M3.1 FlashEffort not named in source tableMiniMax29Evaluations →
25Step 5 PreviewEffort not named in source tableStepFun / 阶跃星辰29Evaluations →
27GPT-6 LunamaxOpenAI28Evaluations →
27GLM-5.3-FlashEffort not named in source tableZ.ai / 智谱28Evaluations →
29MiMo-V2.6-FlashEffort not named in source tableXiaomi MiMo / 小米27Evaluations →
30Mistral Large 4 (high)high · released weights; the source table also lists medium at 23Mistral AI26Evaluations →
31Muse Spark OpenxhighMeta25Evaluations →
32Qwen3.8 27BxhighAlibaba / Qwen20Evaluations →
33Beam 1.0Effort not named in source tableReflection18Evaluations →
34Nemotron 3.5 Super VLEffort not named in source tableNVIDIA14Evaluations →
35Muse GlimmerhighMeta6Evaluations →
How to read this leaderboard

Included models: 35 models, each at its highest tested reasoning setting on the source page; unnamed settings retain the source label. Sorted by that source’s displayed score, not a global top 35. Equal integers do not establish equal unrounded results. Effort and fallback settings differ, so this is not an equal-compute comparison. v5.0/v5.0.1 scores are not comparable with earlier versions.

Rows follow source scores; equal displayed scores share a rank and filtering does not renumber them. Different boards measure different things and are not combined into a total. Works, speed, cost and access still need separate examination.

Read the methodology ↗

How companies reached this generation

Read releases through what came before and after, with the original explanation for each step.

Canva Research ↗

Researches image foundation models and design tools; distinguish image generation from conversion into editable designs.

View 1 evolution entry
  1. Canva Craft

    Originally announced October 22. Craft generation and later Magic Layers conversion have separate interfaces, access conditions and credits; our October 26 analysis is not a second model launch.

    Release source ↗

KRAFTON AI ↗

Develops the Raon family; distinguish preview state, hosted access and local-runtime conditions.

View 1 evolution entry
  1. Raon-1 mini preview

    This is our observation date, not an inferred launch date. The official card documents a 28.8B/4.3B text MoE preview under Apache 2.0; local deployment depends on quantization, runtime branch and hardware.

    Checked source ↗

Nanbeige LLM Lab / 南北阁 ↗

Develops text-reasoning, coding and agent models; keep teachers separate from planned distilled students.

View 1 evolution entry
  1. Nanbeige4.3-33B

    This is our observation date, not a launch inferred from HF creation. The 33B dense teacher and Base have repositories; the planned 3B student remains future work.

    Checked source ↗

Block ↗

A small AI research team within Block explores speech training and live conversation with Dessa; research demonstrations and public products are separate.

View 1 evolution entry
  1. Dessa · 全双工语音研究展示

    The original X announcement is October 23 at 18:23 UTC, October 24 in Beijing. The page presents recordings and an M4 Max claim; general downloads/waitlist remain prospective, without establishing open weights.

    Release source ↗

InclusionAI ↗

Keep Ling open models and hosted evaluation versions distinct, with source-specific weight and configuration information.

View 1 evolution entry
  1. Ling 3.1 Flash 1019 · AA托管推理版本观察

    AA table and model page observed at44. This is the observation date, not a launch date inferred from1019. The open-family model card is separate.

    Release source ↗

Recursive ↗

Explore automated research across models, training and local inference; inspect Flint capability, quantization and hardware separately.

View 1 evolution entry
  1. Flint · 模型与本地引擎公开

    A 39.8B-total, roughly 2.5B-active text coding model and rec engine are released; publisher coding results do not map to the existing AA, Pi or physics boards.

    Release source ↗

Xiaomi MiMo / 小米 ↗

MiMo connects multimodal understanding and long tasks to APIs and coding tools; distinguish published RL weights, served model identities and retirement of older names.

View 2 evolution entries
  1. MiMo-V2.6 Pro / Flash · 正式发布

    Model Release lists the multimodal Pro/Flash models and UltraSpeed serving tier; the Studio changelog separately dates integration September 21.

    Release source ↗
  2. MiMo-V2.5 Pro / V2.5 · 旧名称下线

    Official documentation retires the old IDs with errors at 10:00 Beijing time and asks users to switch to V2.6; this is endpoint retirement, not a V2.6 launch that day.

    Release source ↗

Tencent Hunyuan / 腾讯混元 ↗

Follow specialist Hunyuan models and their roles in applications. Long-term memory writing and simultaneous interpretation are distinct models; frameworks, retrieval, speech synthesis and final output remain separate components.

View 3 evolution entries
  1. Hy-Memory-30B-A3B · 研究与模型资料

    Long-term memory-writing research learns what to retain, update and consolidate within memory frameworks. Its four-harness results use fixed 20% benchmark subsamples and do not establish a general-capability rank.

    Release source ↗
  2. Hy-SimulMT · 同声传译开源

    The 30B-A3B Chinese-English simultaneous speech-translation model opens under Apache-2.0, with FP8/GGUF variants for two bases. Vendor results show ~3-second RealSI lag and 85.8% in the human comparison; they differ from the AA/Pi/Physics board methods.

    Release source ↗
  3. Hy4 正式版 · API 上线,权重承诺开源

    Hy4 official ships on Tencent Cloud and OpenRouter at preview prices (¥6 input / ¥0.3 cached / ¥18 output per million tokens), adding first-time vision and computer use. Weights and system card are promised “in due course” and remain unreleased; third-party boards have no Hy4 row yet. This site read the official research page in full and re-checked the weights repos in the evening round.

    Release source ↗

NVIDIA ↗

Connect coding agents and world simulation through weights, post-training and runtimes; distinguish Nemotron/Cosmos capability, licensing, hardware and actual use.

View 6 evolution entries
  1. Nemotron 3.5 Super VL · 正式发布

    The 120B-total, 12B-active multimodal model launches with public weights, training materials and hosted routes; model cards specify precision and hardware.

    Release source ↗
  2. Nemotron 3.5 Bolt 74B-A9B 24GB · 正式发布

    The 74B-total, 9B-active local-agent model launches. 24GB names a specific quantization recipe; long-context GGUF requires more memory. The October 15 source still described E4B as a later plan; its October 23 release is listed separately.

    Release source ↗
  3. Cosmos3-H-Dreams · 动作条件手术模拟

    A specialist continues endoscopic video from two-arm actions; distinguish two-/four-step weights and the GPU runtime. Clinical validity and a matching score in the existing physics board remain unestablished.

    Release source ↗
  4. Nemotron 3.5 Bolt E4B · BF16与4GB制品公开

    The research and card date GA to October 23: Gemma-4-E4B post-training, 8B total/4B effective, with BF16 and Q4_K_M GGUF artifacts. 4GB describes a target VRAM envelope, not total system memory.

    Release source ↗
  5. Nemotron 3.5 Transcribe 30B-A4B · 开放权重

    A speech-transcription SALM appears on HF trending: 40 languages, recordings to one hour, 34B total / about 4.3B active including the audio encoder, BF16/FP8/NVFP4, servable on one H100, OpenMDW-1.1 license. Card read October 29; the card does not state a weight-release date, so this is our check date.

    Release source ↗
  6. Cosmos 3 · 四个开放权重后训练例

    The official HF blog announces Cosmos3-Nano-Transfer-Auto/Sim-Auto/Sim-Transfer/Sim-Bimanual: 16B trainable parameters, OpenMDW1.1, for 11-view + LiDAR driving simulation and bimanual robot actions to video; repos created August, weights updated Oct 30, blog Nov 5. Physics board gains Cosmos rows, Super (Astra) 65.47 second; AA/Pi do not apply.

    Release source ↗

MiniMax ↗

Follow the formal M3.1 Flash model, effort settings and independent results, keeping API, subscription and later weights separate.

View 4 evolution entries
  1. MiniMax M3.1 Flash

    The formal model reaches API, M Plan and MiniMax Code with 1M context and five effort settings; weights and Pro follow separately.

    Release source ↗
  2. M3.1 Flash · 公开权重文件核对

    Fixed54e4e69d card,48 main shards/DSpark, NVFP4 serving and community license read. This is the check date, not a weight-release date inferred from repository timestamps.

    Release source ↗
  3. MiniMax H3 · 音画同生开放权重核对

    Model card and license read: the H3-Omni-Transformer is a 33B dense single-stream Transformer (~13B AdaLN not loaded at inference), 4–15 s at 24 fps 32 kHz stereo up to 2K; the HF repo was created July 28 (API-checked Nov 9), the license text states release/license date 2026-08-02, and the applicable territory excludes the US, EU, UK and South Korea; the community compression derivative (HyperFlow H3 Flash) ships under the same license and is not an official product.

    Release source ↗
  4. MiniMax H3-Image · 开放权重模型卡核对

    The official image-generation-and-editing model card, read: a 33.1B single-stream DiT on the H3 design minus the audio branch, reusing the MiniMax-H3 text encoder; about 16MP direct output (4096×4096 up to 21:9), distilled to 6 steps with no CFG, and up to 16 reference images for editing; weights include a 66.28GB transformer and a 2.34GB VAE, the text encoder loads from the H3 repository, with a local prompt-expander (PE) alongside. This is the check date; the repository's creation (Nov 3) and the actual availability moment are recorded separately, under the H3-Image Community License. No independent image evaluation carries a matching entry yet — the gap stays stated.

    Release source ↗

PrismML ↗

Bonsai 2 uses ternary weights to shrink local model files; distinguish backbones, vision components, runtimes and actual memory use.

View 2 evolution entries
  1. Ternary Bonsai 2 27B

    The ternary model derived from Qwen3.8-27B launches under Apache 2.0; file size and vendor retention figures are distinct from device testing.

    Release source ↗
  2. Ternary Bonsai 2 2B / 4B / 9B

    Three smaller Qwen3.5-derived models join the family with GGUF and MLX weights; vision components and client integrations have their own runtime conditions.

    Release source ↗

Cactus Compute ↗

Local speech transcription, tool selection and field extraction for small devices; distinguish Whistle and Needle 3 weights from SDK execution.

View 1 evolution entry
  1. Whistle · 原发布

    A 16.9 MB speech model supports seven languages and shares a CPU engine with Needle 3; tool-call JSON still needs software connected to services or devices.

    Release source ↗

TypeSafe AI ↗

System One connects evidence and allowed options to probabilistic decisions; distinguish the 1.14 preview from 1.13, which the stable alias still selects.

View 2 evolution entries
  1. Jev / System One · 首次公开发布

    The first System One model launches; the family-launch date does not establish the later 1.13 release date.

    Release source ↗
  2. Jev 1.14 · 预览可用

    jev-preview selects 1.14 while jev-latest stays on stable 1.13; both default to 64k per request and 32k per question, with an optional longer mode.

    Release source ↗

Odyssey ↗

Connect action-conditioned environment prediction with simulation and physical control; distinguish the public preview from benchmark configurations.

View 2 evolution entries
  1. Odyssey-3 · 技术介绍

    The foundation world-model research describes observations, actions and adaptation to machine policies.

    Release source ↗
  2. Odyssey-3 · 研究预览

    The Flash browser preview opens alongside standard/Pro physical-video evaluations, with distinct configurations.

    Release source ↗

Krea ↗

Models and products for image editing, references and creative iteration; distinguish vendor speed examples from our independent source boards.

View 1 evolution entry
  1. Krea 3 · 图像编辑

    The native-2K editing model accepts up to six references through Krea and partner platforms. Queue and pricing conditions in the launch material are not our task comparison.

    Release source ↗

Elorian ↗

Glimpse revisits images and video intervals during visual reasoning; distinguish its observations and plans from the robot policy that executes actions.

View 1 evolution entry
  1. Glimpse · 研究预览

    The API and chat research preview presents visual reinspection, video understanding and high-level robot planning; its vendor tests and examples are separate from our three independent source boards.

    Release source ↗

StepFun / 阶跃星辰 ↗

Step 5 Preview targets longer tool-driven work, with hosted access and downloadable weights; distinguish the model card, service output limits and independent scores.

View 1 evolution entry
  1. Step 5 Preview · 技术说明

    The technical account describes a 92-layer architecture, long context and agent training; this date establishes the account, rather than the model's first launch or first weights release.

    Release source ↗

Mistral AI ↗

Large 4 weights are open; follow multimodal understanding, coding tasks, license terms and the security positioning.

View 2 evolution entries
  1. Mistral Large 4 · Public Preview

    The next general multimodal model opens an API preview, with weights planned for the end of the month.

    Release source ↗
  2. Mistral Large 4 · 开放权重

    Weights shipped by the month-end promised at preview: Hugging Face hosts an FP8 main repository and an official NVFP4 quantization, under a modified MIT license (running a MaaS business needs a separate commercial license). API pricing is $0.70 per million input tokens and $2.10 per million output tokens.

    Release source ↗

Reflection ↗

Beam advances from preview coding demonstrations to open 1.0 weights; distinguish the text model, hosted API and artifacts made through tools.

View 2 evolution entries
  1. Beam · Preview

    The first model preview targets coding, reasoning and tool tasks, with gradual API access and weights planned for this month.

    Release source ↗
  2. Beam 1.0 · 开放权重

    BF16, FP8 and NVFP4 main weights ship under Apache 2.0 alongside a DFlash2 drafter; the current 1.0 API has 1M combined context.

    Release source ↗

OpenAI ↗

Astra, Sol, Luna and Terra address complex work, everyday tasks, high-volume use and low-cost computer use; the same Luna's Decisions route has separate conditions.

View 9 evolution entries
  1. GPT-6 Astra

    GPT-6's complex-work model connects understanding, tool use and file delivery.

    Release source ↗
  2. GPT-6 Sol / Luna

    Sol and Luna join the GPT-6 family with different capability and cost trade-offs.

    Release source ↗
  3. GPT-6 Sol → GPT-6.1 Sol

    Sol advances to 6.1, with reported gains in coding, computer use and professional work.

    Release source ↗
  4. GPT-6 Luna · Decisions API beta

    The same Luna model gains a dedicated endpoint for predicates, choices and scores; this is API access, not a separate set of model weights.

    Release source ↗
  5. GPT-6.1 Sol · Ultrafast

    The same Sol model gains an API speed tier; its capability version and board identity remain, with separate cost and product-access conditions.

    Release source ↗
  6. GPT-6 Luna · Decisions API GA文档观察

    The formal guide read October 23 marks GA, still using gpt-6-luna. This records the observed documentation state, not a new model launch date. Input, refusals and billing connect to the Jev topic.

    Release source ↗
  7. GPT-6.1 Terra

    Terra advances to 6.1: professional computer tasks score +15 points over GPT-5.6 Terra at about a third of the cost, with standard API prices 50% lower; GitHub Copilot added it the same day. The AA board had no entry for it as of that day.

    Release source ↗
  8. Model Spec · 五价值观重组

    The behavior Spec is reorganized around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency and care. A new “consider context carefully” subsection, an updated chain of command, most detailed style guidance retired, and the model now called “GPT.” This is a Spec update, not a new model.

    Release source ↗
  9. GPT-6 Cyber

    A dedicated cybersecurity model: 48.5% on ExploitGym (from 32.1%), Preparedness cybersecurity Critical (up from High); API and Codex first reach approved Daybreak Red customers, expanding progressively, with GPT-5.6 Cyber opening to all Daybreak Red at the same time. Outside the three boards; the gap stays noted.

    Release source ↗

Anthropic ↗

Distinguish Claude's generally available and restricted research models; examine creation, revisions and research validation alongside benchmarks.

View 7 evolution entries
  1. Claude Mythos 5 · 受限发布

    Shares an underlying model with Fable 5, with different safeguards and access; the initial release is for selected partner organizations, not general availability.

    Release source ↗
  2. Claude Mythos 5 → 5.1

    The later 5.1 continues verified access; the October quantum-circuit research explicitly uses Mythos 5, not 5.1, Fable or Opus.

    Release source ↗
  3. Opus 5 → Opus 5.5

    An update to complex work and communication, with public code-animation and same-prompt game examples.

    Release source ↗
  4. Sonnet 5 → Sonnet 5.5

    The second 5.5-family model emphasizes everyday work, speed and effort-dependent costs.

    Release source ↗
  5. Haiku 4.5 → Haiku 5.5

    The lightweight model gains adjustable effort and lower short-prompt rates for extraction, classification and subagent work.

    Release source ↗
  6. Fable 5.2 与 Mythos 5.2

    One underlying model, two doors: Fable reaches everyone with cybersecurity/biology safeguards (blocked queries fall back to Opus 4.8/5), Mythos drops the layer for vetted organizations via the Cyber and Life Sciences Verification Programs; $8/$40 per million tokens below 5.1, fast mode 2.5×. AA v5.0 independent board leads at max 65 (our full-board re-read, November 10).

    Release source ↗
  7. Claude API effort 档位与对话中换挡

    Official docs confirm output_config.effort with five levels (max/xhigh/high/medium/low; most models default to high, Opus 5.5 and Haiku 5.5 to medium), stable on Fable/Mythos 5.2, 5.1 and 5 among others; changing the top-level setting breaks the prompt cache, while per-message switches via a system message inside messages keep it — one conversation can shift from chat-tier to deep-analysis effort. Read from the official docs November 12; reported the same day by IT之家.

    Release source ↗

Google DeepMind ↗

Follow language, image and retrieval models: long tasks, image edits and on-device media search.

View 10 evolution entries
  1. Gemini 3.7 Flash → 3.8 Flash

    Flash advances coding, multi-step reasoning and agent work at the same speed and price positioning; Flash Cyber is a separate security-focused variant.

    Release source ↗
  2. Gemini 4 Argon

    The new generation is announced with access initially limited to invited testers; evaluation results do not establish public availability.

    Release source ↗
  3. Nano Banana 2.1

    Image generation and conversational editing advance, emphasizing local edits, subject consistency and text rendering.

    Release source ↗
  4. EmbeddingGemma 2

    The retrieval model extends text representations to images, audio and video, with public weights and phone demos.

    Release source ↗
  5. Gemini 4 Argon · 扩大开放

    The same model expands to the paid API, Gemini Pro/Ultra and development tools; shared context and long-decode conditions matter. Copilot has separate plans and gradual rollout.

    Release source ↗
  6. Genie 4

    A new generation of the real-time interactive world model: native audio, on-the-fly text steering and Gemini-planned multi-step events; a research preview reaches trusted testers first, then Project Genie for AI Ultra in coming weeks. No board entries — the scope gap stays stated.

    Release source ↗
  7. Gemini 4 Carbon

    A new frontier model: first on AA v5.0 at 59 (high), officially positioned for cybersecurity and long-horizon work. Access reaches a small group of cyber defenders this week, widens via Fairwind, and stays pending for developers and consumers; introductory pricing is $2/$10 per million tokens, later $4/$20. A high score is not general availability.

    Release source ↗
  8. Gemini 4 Carbon · 普遍开放

    Two weeks after launch, Carbon opens to all paying developers and enterprises via API (AI Studio / Gemini for business, still $2/$10 per million tokens introductory), and to AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search; GitHub Copilot carries it the same day. The post adds new benchmark claims and behavior figures (agentic honesty 92% versus Argon’s 70%; adaptive-red-team per-attempt breach rate 0.8% versus 3.1%) and names AlphaProof Code’s Signal verification as a flagship use.

    Release source ↗
  9. Gemini Robotics 2.5

    The vision-language-action model rolls out to robotics partners: a kitting pilot on two active Mercedes lines reached 96% (proof-of-concept caliber), with success rates 30–100% above the previous baseline and tasks 20–30% faster across 200 tasks; Gemini 4 Argon accelerated training as a simulation engine. No public access; not applicable to the AA / Pi / Physics boards — details in our article.

    Release source ↗
  10. Gemini 4 Hydrogen

    The family adds an everyday-execution tier, available to everyone on day one: all AI Studio tiers, the Gemini API, business surfaces, the Gemini app and Google Search (paid Pro/Ultra also get Antigravity and Sheets); $1.5/$7.5 per million tokens flat across the 1M-token context, with four thinking levels including the new xhigh. The AA independent board lists xhigh 54 / high 51 (above Argon's high 50). Antigravity ships Orchestrator the same day, switching between Carbon and Hydrogen within a session — officially near-full-Carbon success on Terminal-Bench 4.0 with 24% less quota and 19% faster.

    Release source ↗

Meta ↗

Muse Spark, Wonder and Full reach tasks through services; Spark Open separately offers deployable weights. Distinguish capabilities, memory requirements, licensing and product access.

View 6 evolution entries
  1. Muse Spark 1.1 → 1.2

    Coding, debugging and codebase understanding advance through co-training with Muse Code; the model and terminal agent launch together as distinct layers.

    Release source ↗
  2. Muse Glimmer

    The official blog explicitly dates the open roughly 30B dense multimodal Apache 2.0 weights to August 10. Cataloguing it and AA/high 17 on October 26 does not change that event date.

    Release source ↗
  3. Muse Spark 1.2 → 1.3

    The release emphasizes longer tasks, tool use and coding, with max effort available in Muse Code and Meta Model API.

    Release source ↗
  4. Muse Wonder / Muse Wonder Full · 首次发布

    The new series emphasizes Active Memory, Active Focus, long-running tasks and coding. Wonder opens through Meta Model API and Muse Code; Full is for Muse Code subscribers. Personal Muse integration remains a later rollout.

    Release source ↗
  5. Muse Spark Open · 权重发布

    The 816B-total, 42B-active multimodal model releases weights under the Muse Model License. It derives from August’s Spark with additional safety post-training, rather than the same weights as Spark 1.3 or Wonder.

    Release source ↗
  6. Muse Ember · 首次发布

    Meta’s smallest and cheapest model yet, built for the volume work agents repeat all day — tool calls, coding subtasks, routing, extraction and long-document parsing; natively multimodal inputs (text, image, audio, video). Meta Model API pricing is $0.10 input / $0.50 output per million tokens (up to 200k input tokens; $0.30/$1.50 beyond); the official page reports 36 on the AA intelligence index. Meta pairs it with Wonder or Spark: Wonder plans, Ember handles the volume.

    Release source ↗

SpaceXAI ↗

Follow Grok's coding and knowledge-work updates, separating standard models from faster serving tiers.

View 2 evolution entries
  1. Grok 4.5 → 4.6

    Further training advances long tasks, coding and interactive work through Grok Build, Cursor and the API; faster serving has separate pricing.

    Release source ↗
  2. Grok 4.6 → 4.7

    A larger base model emphasizes long tasks and self-checking, at the same standard token rates as 4.6.

    Release source ↗

Alibaba / Qwen ↗

Distinguish text Max, Flash-Next and 27B from image 2.1/Turbo; capabilities, weights terms and hosted routes have specific conditions.

View 6 evolution entries
  1. Qwen3.8 Max → Max 0902

    The 0902 snapshot updates coding, tool orchestration and vision while retaining explicit version identity.

    Release source ↗
  2. qwen3.8-max → Max 0902 (default)

    The announced default Max endpoint transition points to 0902; this is a routing change, not another new model.

    Release source ↗
  3. Qwen-Image-2.1

    Generation, transparent assets and multi-reference edits share one visual model; check the terms for available weights.

    Release source ↗
  4. Qwen-Image-2.1 Pro API

    The Singapore International lifecycle table lists hosted Pro access; the repository's October 9 link announcement is not its initial release.

    Release source ↗
  5. Qwen-Image-2.1 → Turbo

    An eight-step checkpoint and hosted Turbo access continue multi-reference editing; denoising steps are not measured end-to-end latency.

    Release source ↗
  6. Qwen4-Max

    Qwen4's first flagship ships with API access: a limited qwen4-max-preview opened November 11; about a 2.4T sparse-MoE backbone, native vision and 1M context, with weights not open and smaller open models to follow; the AA board shows a 43 row the same day.

    Release source ↗

DeepSeek ↗

Flash's vision and tool capabilities continue through the 1017 project-development update; distinguish weights, API routing and search/fetch services.

View 7 evolution entries
  1. V4 Pro / V4 Flash

    The V4 family reaches the API with Pro and Flash offering different capability and resource trade-offs; older request names temporarily map to Flash.

    Release source ↗
  2. V4 Flash Preview → Flash 0731

    The architecture and size stay unchanged while post-training and Responses API support advance; this update applies only to the Flash API, not Pro or the web model.

    Release source ↗
  3. V4 Pro · GA

    Pro reaches general availability in the app, web and API with Responses API support and more flexible thinking effort; its request name stays unchanged.

    Release source ↗
  4. V4 Flash → Flash Vision Exp

    A separate experimental vision model adds image understanding to Flash; V4.1 Flash later succeeds it.

    Release source ↗
  5. V4 Flash / Vision Exp → V4.1 Flash

    Native multimodal Flash is released; legacy Flash API names temporarily route to the new model while Pro remains available.

    Release source ↗
  6. V4.1 Flash → Flash 1017

    A minor update ships different weights across website/app/API under the same deepseek-flash name and model prices; separate search/fetch APIs charge by request.

    Release source ↗
  7. Harness 公开预览 · Super-agent 研究原型

    Harness enters worldwide public preview and open source (everything-is-a-plugin on the Cordis microkernel, Web UI via npx); the same day it open-sources the dsh-superagent-prototype (MIT) research prototype: a proactive coding agent listening to GitHub/Feishu, with Router answering people while Workers execute in isolation and a panel listing only what needs a person. The README states the prototype differs from the internal version and the design article's implementation; experiments are vendor synthetic-load figures. Agent infrastructure, not a new model; repository docs read November 1.

    Release source ↗

Z.ai / 智谱 ↗

GLM's flagship evolves around long tasks and software engineering, while Flash adds vision separately; keep models, chat products and coding agents distinct.

View 5 evolution entries
  1. GLM-5

    The GLM-5 generation advances systems engineering and long-running agent tasks.

    Release source ↗
  2. GLM-5 → GLM-5.1

    The update advances engineering planning, terminal tasks and sustained execution; outcomes still depend on the task and tool configuration.

    Release source ↗
  3. GLM-5.1 → GLM-5.2

    A 1M-token context and selectable reasoning effort accompany further long-task and coding improvements.

    Release source ↗
  4. GLM-5.2 → GLM-5.3

    Post-training advances complex engineering and agent tasks on the GLM-5.2 base; the flagship still takes text input.

    Release source ↗
  5. GLM-5.3-Flash

    A separate new base brings native vision to coding, browser and document work, with API and open-weight routes.

    Release source ↗

Moonshot AI / 月之暗面 ↗

K3's flagship path and Code's efficiency-focused models evolve alongside each other; follow vision, long tasks and the model behind each endpoint.

View 5 evolution entries
  1. Kimi K2.6

    Long-horizon coding and agent work advance with open weights, providing a previous-generation reference for later general models.

    Release source ↗
  2. Kimi K2.7 Code

    The coding branch launches with open weights, exploring longer engineering work and reasoning-token trade-offs.

    Release source ↗
  3. Kimi K3

    Native vision and 1M context reach long-horizon coding and knowledge work; the full weights follow on July 27.

    Release source ↗
  4. Kimi Code → K2.8 Preview

    The standard kimi-for-coding endpoint moves to K2.8 Preview with more efficient reasoning, alongside the K3 flagship.

    Release source ↗
  5. Kimi K3.1 Pro / Highspeed · Code上线

    Code records rollout on October 23, followed by the October 24 research-index entry. Separate 1M eligibility, thinking and quotas for Pro/Highspeed; the weights were promised “by November 4” — this site's Nov 3/4 morning checks found nothing, public evidence starts with the first community mirror at 23:59 Beijing on Nov 4, and a Nov 5 re-check confirms 96 shards ≈1.56TB are public. Honored at the wire; see the site article for the reading.

    Release source ↗

Cloudflare ↗

Bounded decisions let software take the next step; distinguish weights, hosted windows and access costs.

View 2 evolution entries
  1. Clef / Clef-flash

    Two decision models return probabilities over defined options, illustrated with message classification and web workflows.

    Release source ↗
  2. Clef-omni / Clef-flash托管变化

    Audio/video joins bounded decisions; Flash hosted input rates fall while its advertised 64k window becomes 24k. Published Flash weights are unchanged.

    Release source ↗

Liquid AI ↗

A small open-weight family for on-device use: one family spans text, vision, audio and omni; hardware conditions and specialist capabilities are checked separately, and vendor benchmarks are not independent evidence.

View 1 evolution entry
  1. LFM3 家族 · 发布

    The LFM3 open-weight family launches (Base/1.7B text with 256K context, VL-2B, Audio-2B, Omni-2B and d1 decision variants), pitched at on-device deployment; Omni-2B is ~2.2B parameters taking image/video/audio/text in and text plus speech out. Card benchmarks and its Oct 31 HF trend popularity are recorded as vendor figures and popularity facts, not capability evidence.

    Release source ↗

From scores to evaluations

Continue through each model’s sources, test conditions and related works.

Canva Research

Canva Craft

Canva Research’s image foundation model plans composition before rendering up to 2K images. The generation task returns an image and optional asset; Magic Layers conversion into an editable design is a separate task and credit use. Publisher language/rank claims are not our Chinese-text tests.

Evaluations & evolution →

KRAFTON AI

Raon-1 mini preview

A text-only 28.8B-total/4.3B-active MoE preview with native 262,144-token context and Apache 2.0 weights. Official 32GB NVFP4 use differs from MXFP8 offload; llama.cpp upstream support is pending. Publisher evaluations do not fill our three boards, and we have not run it.

Evaluations & evolution →

Nanbeige LLM Lab / 南北阁

Nanbeige4.3-33B

A 33B dense text teacher with 512K context and tool use for coding and cowork, under Apache 2.0. The planned 3B student is not this release. Publisher AA-LCR/Terminal-Bench results and simulated benchmark dollars are not our intelligence/Pi scores or revenue; we have not run it.

Evaluations & evolution →

Meta

Muse Glimmer

Meta’s roughly 30B dense open multimodal model accepts text/images and outputs text under Apache 2.0. The AA text index reads 6 (high) under the v5.0 caliber; the v4.3.2-era 17 stays in records — the two versions do not mix, and this site's Nov 16 recheck confirms the source row still displays 6. Quantized 24/32GB target configurations differ from full precision; this August release is not today’s launch, and text-index results do not establish all agent performance.

Evaluations & evolution →

Block

Dessa

Block’s full-duplex speech research demonstration claims audio-only training and local inference on MacBook Pro M4 Max or better. Download and waitlist remain prospective; the page does not specify public weights, model licensing or API terms, and the three boards have no comparable result.

Evaluations & evolution →

NVIDIA

Nemotron 3.5 Bolt E4B

A text model post-trained from Gemma-4-E4B, with 8B total/4B effective parameters and BF16/Q4_K_M GGUF artifacts. 4GB names a target VRAM envelope; file size, cache and system memory differ. AA has no numeric result, Pi has no matching result, and Physics video does not apply.

Evaluations & evolution →

NVIDIA

Nemotron 3.5 Transcribe 30B-A4B

An open-weight speech-transcription SALM: 40 languages, recordings up to one hour, a 65,536-token context, 34B total / about 4.3B active including the audio encoder, BF16/FP8/NVFP4, servable on one H100 80GB under OpenMDW-1.1. Transcription is a specialist capability — the AA/Pi text indices and the Physics video board do not apply; the card's own benchmarks (multi-speaker cpWER 15.80, English WER 4.71 and more, captured October 12) read as vendor figures.

Evaluations & evolution →

Liquid AI

LFM3-Omni-2B

Liquid AI's first full-modality model: about 2.2B parameters, taking image/video/audio/text in and text plus 24kHz speech out, 32K context per the model card, 16 text and 19 speech languages, LFM Open License v1.0, non-reasoning. The official family also spans Base/1.7B text (256K), VL-2B, Audio-2B and d1 decision variants. On-device full-modality is a specialist capability — the AA/Pi text indices and the Physics video board do not apply; card benchmarks (Daily-Omni 70.1 and more) and its Oct 31 HF trend popularity are recorded separately as vendor figures and popularity facts.

Evaluations & evolution →

Moonshot AI / 月之暗面

Kimi K3.1 Pro

A post-trained K3 model: Code rollout is dated October 23, the research index October 24. AA text index 55 is the v4.3.2 figure of October 23–24; after the v5.0 board switch the source row reads 46 (effort still unnamed; v4.3.2 and v5.0 scores are not mixed). Distinguish Highspeed access from weights: the full weights promised “by November 4” are now public at the wire — this site's Nov 5 re-check confirms 96 shards ≈1.56TB (MXFP4, single commit Nov 2); until now the model was reachable only via the OpenRouter/official APIs.

Evaluations & evolution →

InclusionAI

Ling 3.1 Flash 1019

A hosted reasoning version separately listed by AA at44, with no speed result. AA marks it proprietary; the MIT card for the similarly named open family does not establish identical weights or an inference recipe.

Evaluations & evolution →

NVIDIA

Cosmos3-H-Dreams

A surgical world model generates video from an initial endoscopic frame and 28-dimensional two-arm actions for robotics simulation research. Model terms differ from Apache2 code. Training 10Hz, inference FPS and clinical validity are distinct; no generic Cosmos or Odyssey physics score is inherited.

Evaluations & evolution →

NVIDIA

Cosmos 3 Super

The GA base of the Cosmos 3 family (the largest of Super/Nano/Edge): open-weight HF repo created March 10, updated September 16, 38k downloads at observation. Physics-IQ rows Cosmos 3 Super (Astra) 65.47 (Astra prompt upsampling + BoN×8) and single-run 50.80 are board calibers; AA/Pi target language and coding models and do not apply.

Evaluations & evolution →

NVIDIA

Cosmos 3 Nano

The 16B-trainable base of the Cosmos 3 family: 137k downloads on the GA repo at observation; four open-weight post-trained checkpoints (Transfer-Auto/Sim-Auto/Sim-Transfer/Sim-Bimanual) target 11-view + LiDAR driving simulation and bimanual robot actions to video, OpenMDW1.1, 32.4–34.5 GB per repo, vendor-reported wins over 14B teachers. Physics rows 61.83 (Astra + BoN×8) and single-run 43.00; AA/Pi do not apply.

Evaluations & evolution →

Recursive

Recursive Flint

A 39.8B-total, roughly 2.5B-active local text coding model with Apache2 weights. rec requires supported hardware and at least 48GB unified memory. Publisher Terminal-Bench2.1/SWE results do not match the three existing boards.

Evaluations & evolution →

Xiaomi MiMo / 小米

MiMo-V2.6 Pro

Flagship multimodal reasoning with 1M context; the published RL card lists 1.02T total, 42B active parameters and MIT licensing. AA records 46 in the same table; weights, hosted API and UltraSpeed serving are distinct runtime conditions.

Evaluations & evolution →

Xiaomi MiMo / 小米

MiMo-V2.6 Flash

An efficiency-oriented multimodal model with 1M context; the published RL card lists 309B total, 15B active parameters and MIT licensing. Its AA score is 38, not Pro’s 46; Flash does not guarantee consumer-device deployment.

Evaluations & evolution →

Tencent Hunyuan / 腾讯混元

Hy-Memory-30B-A3B

A specialist long-term-memory writer with about 30B total and 3B active parameters, organizing and updating memory inside external frameworks. Retrieval and final answers remain separate. Vendor comparisons on fixed 20% subsamples across four harnesses differ from AA text intelligence, Pi coding and Physics video continuation; no same-identity score currently qualifies it for those boards.

Evaluations & evolution →

Tencent Hunyuan / 腾讯混元

Hy-SimulMT (30B-A3B)

A specialist bidirectional Chinese-English simultaneous speech-translation model with about 30B total and 3B active parameters, open-sourced under Apache-2.0 (third-party components keep their terms) alongside Interleave and FP8/GGUF variants. Vendor results (~3-second RealSI lag, 85.8% against human interpreters) use GEMBA judged by GPT-6.1-Sol; speech output needs a separate TTS. Simultaneous interpreting differs from AA text intelligence, Pi coding and Physics video continuation, so no same-identity score currently qualifies it for those boards.

Evaluations & evolution →

Tencent Hunyuan / 腾讯混元

Hy4

The official version of Tencent Hunyuan's newest flagship: the API is live on Tencent Cloud and OpenRouter at preview prices (¥6 input / ¥0.3 cached / ¥18 output per million tokens), with first-time vision and computer use; officially “the strongest open models” ranking entry, yet the weights and system card remain under a due-course promise as of 2026-11-04 (only preview weights on GitHub/HF, re-checked here). On the November 5 full AA re-read, Hy4 joins the source table at 35 (reasoning version, effort not named in the source table). “Strongest open models” remains an official characterization; appendix and audit figures are official claims; nothing reproduced here.

Evaluations & evolution →

NVIDIA

NVIDIA Nemotron 3.5 Super VL

120B total and 12B active parameters, up to 1M context, with public weights under OpenMDW 1.1. The BF16 card gives 4×H100/2×GB200 examples; NVFP4 has separate conditions. AA26 does not name effort or establish customer outcomes after specialized post-training.

Evaluations & evolution →

NVIDIA

Nemotron 3.5 Bolt 74B-A9B 24GB

74B total and 9B active parameters, up to 256K context, OpenMDW 1.1. The GGUF card lists 28GB/65,536 and 32GB/262,144-token single-sequence recipes; check vision components, drafter and hardware together. Not installed or tested here.

Evaluations & evolution →

MiniMax

MiniMax M3.1 Flash

API access and five effort settings continue. Public weight shards, 428B total/23B active and the community license were checked October 24. Retain AA 43 with unnamed effort; serving recipes are not local performance tests.

Evaluations & evolution →

MiniMax

MiniMax H3

An open-weights joint audio-video model: the official card puts the H3-Omni-Transformer at 33B dense single-stream (~13B AdaLN not loaded at inference), generating video with native 32 kHz stereo audio up to 15 seconds, 24 fps, up to 2K, plus an H3-Context-IR input-understanding module. The community license applies outside the US, EU, UK and South Korea. AA / Pi are text boards and do not apply; Physics-IQ scores video continuation (v2v), a different track from text- or image-to-video, and its rows read on 2026-11-07 contain no H3 — with no same-caliber score it stays off the boards and unrated.

Evaluations & evolution →

PrismML

Ternary Bonsai 2 2B

A Qwen3.5-2B-derived vision-language model. PTQ1_0 language weights are about 0.43GB, not peak RAM; check vision and runtimes separately. No matching AA/Pi result is inherited from the backbone; Physics video continuation does not apply.

Evaluations & evolution →

PrismML

Ternary Bonsai 2 4B

A Qwen3.5-4B-derived vision-language model. PTQ1_0 language weights are about 0.95GB, not peak RAM; check vision and runtimes separately. No matching AA/Pi result is inherited from the backbone; Physics video continuation does not apply.

Evaluations & evolution →

PrismML

Ternary Bonsai 2 9B

A Qwen3.5-9B-derived vision-language model. PTQ1_0 language weights are about 2.00GB, not peak RAM; check vision and runtimes separately. No matching AA/Pi result is inherited from the backbone; Physics video continuation does not apply.

Evaluations & evolution →

PrismML

Ternary Bonsai 2 27B

The earlier 27B Bonsai 2 model derives from Qwen3.8-27B. Keep its specifications and tests separate from the later Qwen3.5-based models. These ternary weights do not inherit backbone scores; no matching AA/Pi result, and Physics video continuation does not apply.

Evaluations & evolution →

TypeSafe AI

Jev 1.14 Preview

A bounded decision model for text and structured text. jev-preview selects 1.14; stable jev-latest remains on 1.13. Defaults: 64k per request and 32k per question, with optional long context. No comparable result in our three boards; earlier game results concern another version.

Evaluations & evolution →

Meta

Muse Spark Open

816B total parameters and 42B active, with text, image, audio and video inputs and text output. Self-host under the Muse Model License; published quantized examples still occupy hundreds of GB of memory. AA at xhigh scores 39, separately from Spark 1.3 and Wonder.

Evaluations & evolution →

Meta

Muse Wonder Full

The Full model in the Wonder series emphasizes memory, sustained execution and coding. Its AA max-setting text index is 56, separate from standard Wonder. Current access is through a Muse Code subscription; this score does not describe the standard API model.

Evaluations & evolution →

Meta

Muse Wonder

The new series targets long-running tasks, computer use and coding, with Active Memory and Active Focus describing context and sustained work. Its AA max-setting text index is 52. Meta Model API and Muse Code access are open; personal Muse integration is a later rollout, and Full is a separate model.

Evaluations & evolution →

Meta

Muse Ember

Meta’s smallest and cheapest model yet, positioned for the volume work agents repeat all day — tool calls, coding subtasks, routing, extraction and long-document parsing; natively multimodal inputs (text, image, audio, video). Meta Model API pricing is $0.10 input / $0.50 output per million tokens (up to 200k input tokens; $0.30/$1.50 beyond); its AA intelligence index is 36 at max (34 at xhigh), with the official page reporting about $0.13 per task at max reasoning. Official benchmarks are strongest on automation and terminal execution, while FrontierCode 2.0-AA at 27.2 still trails Haiku 5.5’s 34.9; Meta pairs it with Wonder or Spark in a plan-then-execute split. Access runs through Muse Code subscriptions and the Meta Model API (model name muse-ember-1.0); the official page carries no date, so the release is recorded as November 9 per same-day coverage — this site read the official page on November 10.

Evaluations & evolution →

Cactus Compute

Cactus Whistle

A 16.9 MB CPU speech model with seven supported languages, excluding Chinese. It can share an engine with Needle 3 to return text and tool arguments. Speech error rates have a separate scope from our text, coding and physics boards.

Evaluations & evolution →

Cactus Compute

Needle 3

A laddered small model with its own weights fills names and arguments for declared tools or schemas; SDK function execution is a later step. Our three boards have no comparable score, and its individual release date is not established by the official material.

Evaluations & evolution →

TypeSafe AI

Jev 1.13

Bounded decisions from text or JSON state. Stable jev-latest still selects 1.13, with 1.14 in preview. Existing game results concern 1.13, not the newer version; its individual release date remains unestablished.

Evaluations & evolution →

OpenAI

GPT-6 Luna

A GPT-6 model for focused, high-volume tasks. Its max-setting text index and dedicated Decisions endpoint have separate scopes, prices and output types.

Evaluations & evolution →

Cloudflare

Clef

Bounded decisions from text, JSON, images or video on a 27B source architecture. Hosted limits and weights differ; no comparable score appears in our three boards.

Evaluations & evolution →

Cloudflare

Clef-flash

A decision model on a 9B source architecture. October 9 changes its hosted price/window, not the weights. No comparable score is in our three boards.

Evaluations & evolution →

Cloudflare

Clef-omni

A 30B-A3B source architecture scores options from audio and video. One scoring call is not a continuous agent. No comparable score is in our three boards; keep the model card distinct from service terms.

Evaluations & evolution →

Mistral AI

Mistral Large 4

A trillion-parameter multimodal model; weights opened as promised on October 29. Read the license, index score and security positioning separately — running it as a service needs a separate license.

Evaluations & evolution →

Reflection

Beam Preview

The October 5 preview and October 6 record of a 256K beta connect to the astronaut-game demo. Formal 1.0 weights opened October 16; checkpoint identity is unconfirmed, so this preview does not inherit the later score.

Evaluations & evolution →

Reflection

Beam 1.0

The open 501B/23B text model has Apache 2.0 main weights and a current API with 1M combined context and 128K maximum output. AA lists 30 without naming effort; keep it separate from the old preview identity and vendor scores.

Evaluations & evolution →

Krea

Krea 3

An image-editing model: six references in the launch, eight in current API docs, with a 1536px longest-side limit from five references. App compute units and beta-workspace API access differ. Base 1K/2K rates are $0.03/$0.06, with reference and high-mode fees added. Sub-eight-second edits are a vendor claim; earlier platform jobs including queues are not our test.

Evaluations & evolution →

Elorian

Glimpse

A visual-reasoning research preview offered through an API and chat. Listed rates are $0.60 per million input tokens and $3 per million output tokens; repeated inspection also takes time and tokens. In the robot example, Glimpse plans and π0.5 executes; this is not direct action generation or our test.

Evaluations & evolution →

StepFun / 阶跃星辰

Step 5 Preview

A 600B-total, 27B-active multimodal preview with hosted access and weights under StepFun Model License 1.0. Hosted docs specify 1M context and 64k maximum output; the model card's output ceiling has separate runtime conditions.

Evaluations & evolution →

Anthropic

Claude Mythos 5

The restricted model used in the quantum-circuit study. Results come from a model, coding environment and author-built evaluator loop, not Opus 5.5 or measured quantum-hardware acceleration. Current family access guidance covers the later 5.1, not general access to this older version.

Evaluations & evolution →

Anthropic

Claude Fable 5.2

Anthropic's strongest general model, first on the independent AA v5.0 board at 65 (max); one underlying model with Mythos 5.2 — Fable reaches everyone with cybersecurity/biology safeguards (blocked queries fall back to Opus), Mythos drops the layer for vetted organizations. $8/$40 per million tokens below 5.1, fast mode 2.5×; vendor-caliber FrontierCode 2.0 xhigh 49.1%, by their count the highest of any released model.

Evaluations & evolution →

Google DeepMind

Gemini 4 Carbon

Google's strongest model, announced October 28 (Pacific time): 60 (high) on AA v5.0 as of the November 4 full-board reread — it led alone until Fable 5.2 passed it — officially positioned for cybersecurity and long-horizon work. SEC-bench Pro 88.2%, ExploitGym 54.8% and CWE-bench 71.0% are vendor figures, with external benchmarks noted as unguarded. Cyber defenders got access first on October 28; general availability arrived November 10 — paying developers and enterprises via API, AI Ultra subscribers via Antigravity, the Gemini app and AI Mode in Search (GitHub Copilot the same day). Introductory pricing is $2/$10 per million tokens, later $4/$20.

Evaluations & evolution →

Google DeepMind

Gemini 4 Hydrogen

Google's everyday-execution tier for the Gemini 4 family: while Carbon and Argon handle deep long-horizon reasoning, Hydrogen targets fast coding and agentic workflows, available to everyone on day one (all AI Studio tiers, the Gemini API, business surfaces, the Gemini app and Google Search; paid Pro/Ultra subscribers also get it in Antigravity and Sheets). Pricing is $1.5 input / $7.5 output per million tokens, flat across the 1M-token context, with 1M-token input, up to 1M-token output, and four thinking levels (low / medium / high default / new xhigh). The AA independent board lists four tiers — xhigh 54 / high 51 / medium 41 / low 27 (high above Argon's high 50) — with the high tier about $1.87 per task; Google's own comparison table trades wins and losses against Sonnet 5.5 and Terra (trailing on Terminal-Bench 4.1, OSWorld-2.1 and Terminal-Bench-Science). Antigravity ships Orchestrator, switching between Carbon and Hydrogen within a session: near-full-Carbon success on Terminal-Bench 4.0 with 24% less quota and 19% faster (vendor figures). This site read the official blog and AA's board via direct fetch on November 13.

Evaluations & evolution →

Google DeepMind

Genie 4

The first real-time interactive world model to generate audio natively: walk with the keyboard, rewrite scenes by text, with Gemini 3.8 Flash as the guiding agent planning multi-step events; scenes hold together about two minutes, double Genie 3. A research preview reaches trusted testers first, then Project Genie for AI Ultra in coming weeks. No AA / Pi / Physics-IQ listing — world models sit outside those boards, so no ranking follows.

Evaluations & evolution →

Google DeepMind

Gemini Robotics 2.5

A vision-language-action model: official evaluations span 200 tasks with success rates 30–100% above the previous baseline (relative values) and tasks 20–30% faster; a kitting pilot on two active Mercedes lines reached 96% (proof-of-concept caliber), with Gemini Robotics ER tracking the workflow and this model controlling hands and arms. Safety adds inference-time guidance and the open-source ASIMOV-Jailbreak benchmark (Argon 99.5% versus 82% for the previous ER 2). Training was accelerated by Gemini 4 Argon acting as a simulation engine. Partner rollout only, no public access; not applicable to the AA / Pi / Physics boards — VLA task success sits outside those calibers, and no third-party benchmark exists yet.

Evaluations & evolution →

OpenAI

GPT-6.1 Terra

A fast, efficient computer-use model: vendor results show +15 points over GPT-5.6 Terra at about a third of the cost, matching Astra at roughly a tenth of the token price; standard API prices drop 50% and GitHub Copilot added it the same day. As of October 27 the AA board has no entry for it (the older GPT-5.6 Terra max sits at 42), so it joins the board only when a same-method score exists.

Evaluations & evolution →

OpenAI

GPT-6 Cyber

A dedicated cybersecurity model: finding and validating exploitable vulnerabilities within researcher-defined scope, covering exploit development, vulnerability research, code review and patching. 48.5% on ExploitGym (from 32.1%), officially Critical for cybersecurity under the Preparedness Framework (up from High); access starts with approved Daybreak Red customers, expanding progressively. Not applicable to the AA / Pi / Physics-IQ boards - specialized cyber capability sits outside those calibers, and the approval-only gate leaves no independent third-party readings yet; the gap stays stated.

Evaluations & evolution →

Meta

Muse Spark 1.3

A historical capability model behind Muse. AA marks it deprecated and removes it from the current table as checked October 22; 48 is a historical reference. This does not establish a Meta service shutdown.

Evaluations & evolution →

Alibaba / Qwen

Qwen-Image-2.1

Generation, transparent assets and multi-reference editing. Image-preference evaluation is distinct from the text, coding and physics boards. Weights are limited to non-commercial research/evaluation; commercial use needs separate permission.

Evaluations & evolution →

Alibaba / Qwen

Qwen-Image-2.1 Turbo

An eight-step checkpoint on the base 2.1 visual architecture, with regional hosted access. No separate comparable Turbo image evaluation was found; base-model scores do not transfer.

Evaluations & evolution →

Alibaba / Qwen

Qwen4 Max

Qwen4's first flagship: 2.4T trunk parameters, 71B active per token, plus 328B N-gram Embedding parameters, with native vision input and 1M context. This site read the official blog body in full via browser render on the morning of November 17 (on November 16 only the title and release facts came through, and figures were first recorded from IT之家's relay): the three architecture items are GDN+QSA (three of every four layers are Gated DeltaNet compressing history into a fixed state, one is Qwen Sparse Attention picking context block by block), Gated Residual's four gated read/write branches, and a layer-two N-gram lookup; the vendor reports spike-free pretraining at 2.4T with base-model BPB about 5–13% below Qwen3.8-Max-Base at roughly 25% fewer active parameters per token. Capability and cost are vendor-reported: Work360 52.69 versus the predecessor's 47.74; token consumption down about 24%/35% (Xhigh/High) on Work360 and about 48%/49% on Terminal-Bench 4.0; API prices cut across the board with cache-hit pricing lowered the most, and Work360 per-task average cost 26.58→11.16 yuan (−58%) — the circulating "over 60% / 62%" relay wording does not match the official 58%, and the official figure stands. The official 16-panel chart is a clean sweep over the predecessor but tops only four panels against other models (Claude Opus 5.5 is higher on most knowledge-work panels). Weights are not open, with smaller series models promised for later open-sourcing; access runs through the Qwen platform (Token Plan or API) plus Qwen Work, Qoder CN, the Qwen app and Qwen Studio, after a limited qwen4-max-preview from November 11. The AA v5.0.1 source table showed a Qwen4 Max row at 43 on our November 16 evening read (effort not named), unchanged on the November 17 morning re-read; the 43 and the self-reported Work360 figures do not substitute for each other.

Evaluations & evolution →

DeepSeek

DeepSeek V4.1 Flash

Retain September vision, file and effort references. AA/max was recorded at 39 on October 19; the October 20 source no longer lists the older name, so it leaves the current board. Its dated comparison remains in the topic; deepseek-flash now serves 1017.

Evaluations & evolution →

DeepSeek

DeepSeek V4.1 Flash 1017

A new MIT-licensed V4.1 Flash weight snapshot keeps the configuration, image/text input and text output. The API retains deepseek-flash and its prices. Independent AA/max scores 45 and Pi/high scores 100% within the small suite; older AA39 and LiveBench results remain with the earlier model.

Evaluations & evolution →

Z.ai / 智谱

GLM-5.3

A text-input model for longer work, advancing complex engineering through post-training on the GLM-5.2 base; explore independent evaluations and a walkable 3D creation.

Evaluations & evolution →

Z.ai / 智谱

GLM-5.3-Flash

Text and images connect to coding, browser and document work; this separate new base offers API and open-weight routes for exploring concrete uses.

Evaluations & evolution →

Moonshot AI / 月之暗面

Kimi K3

Native vision and long context support coding and knowledge work, with an SVG and image description making capabilities concrete; K3 and Code's K2.8 Preview are parallel paths.

Evaluations & evolution →