# Qwen3.8 Max

Alibaba · Open Weight · rank 35 · bench-align-v5

> Every figure below is reproduced as its upstream source published it: nothing is modelled, estimated, interpolated or converted. `Unavailable` means no source published the value — it is never a zero. Each value carries its evidence state and the date it was observed.

Page: https://modelchoice.dev/models/qwen3-8-max  
JSON: https://modelchoice.dev/api/model/qwen3-8-max

## Facts

| Field | Value | Evidence |
| --- | --- | --- |
| Canonical id | `qwen3-8-max` | — |
| Overall score | 65.5 | Observed 2026-11-21 · source benchlm:models |
| Verified score (BenchLM): a separate BenchLM figure, not the overall score | 74 | Observed 2026-11-21 · source benchlm:models |
| Context window | 1M tokens | Observed 2026-11-21 · source benchlm:models |
| Max output (models.dev) | 131,072 tokens | Observed 2026-11-21 · source modelsdev:catalog |
| Modalities (models.dev) | Input: text, image, PDF, video · Output: text | Observed 2026-11-21 · source modelsdev:catalog |
| Release date | 2026-08-03 | Observed 2026-11-21 · source benchlm:models |
| Access type | Open Weight | — |
| Blended $/1M (75% input / 25% output) | Unavailable | Derived from the input and output rates below |

## Capability evidence

Seven axes from the ranking source. An axis the source did not score is unavailable, not zero. Beside each score is BenchLM's separate verified score for that axis; the two are published side by side and are not interchangeable.

| Axis | Score | Evidence | Verified (BenchLM) | Verified evidence |
| --- | --- | --- | --- | --- |
| Agentic | 83.8 | Observed 2026-11-21 · source benchlm:models | 61.3 | Observed 2026-11-21 · source benchlm:models |
| Coding | 70.5 | Observed 2026-11-21 · source benchlm:models | 70.4 | Observed 2026-11-21 · source benchlm:models |
| Knowledge | 70.6 | Observed 2026-11-21 · source benchlm:models | 69.1 | Observed 2026-11-21 · source benchlm:models |
| Reasoning | 79.9 | Observed 2026-11-21 · source benchlm:models | 79.9 | Observed 2026-11-21 · source benchlm:models |
| Multimodal & Grounded | 91.6 | Observed 2026-11-21 · source benchlm:models | 91.6 | Observed 2026-11-21 · source benchlm:models |
| Instruction Following | 95.5 | Observed 2026-11-21 · source benchlm:models | 95.5 | Observed 2026-11-21 · source benchlm:models |
| Math | 72.6 | Observed 2026-11-21 · source benchlm:models | 72.6 | Observed 2026-11-21 · source benchlm:models |

## Runtime service evidence

Measured values with the date they were observed. Nothing is inferred from a sibling model or a provider claim. Regional or per-endpoint measurements appear only when the API supplies them; none are modelled.

| Measurement | Value | Observed | Last good | Evidence |
| --- | --- | --- | --- | --- |
| Time to first token | 52.16 s | 2026-11-21 | 2026-11-21 | Observed 2026-11-21 · source benchlm:speed |
| Throughput | 41 tok/s | 2026-11-21 | 2026-11-21 | Observed 2026-11-21 · source benchlm:speed |

## Endpoint and price matrix

Every published price component, including cache reads and writes.

| Component | USD | Evidence |
| --- | --- | --- |
| Input / 1M tokens | Unavailable | Unavailable · source benchlm:pricing |
| Output / 1M tokens | Unavailable | Unavailable · source benchlm:pricing |
| Cache read / 1M tokens | Unavailable | Unavailable · source benchlm:pricing |
| Cache write / 1M tokens | Unavailable | Unavailable · source openrouter:pricing |
| Blended / 1M (75% input / 25% output) | Unavailable | Derived — from the input and output rates above; it has no source record of its own |
| Cost per successful task (LiveBench) | Unavailable | Unavailable · source livebench:table |

**Pricing note.** Qwen publishes the Qwen3.8-2.4T-A95B checkpoint for self-hosting under the custom Qwen3.8-Max License. Alibaba Cloud Model Studio's official pricing page separately lists the hosted qwen3.8-max SKU in both non-thinking and thinking modes for the 0 < Token <= 1M tier. The China and global tables list CNY 12 input / CNY 36 output per million tokens; the US international table lists CNY 14.988 input / CNY 44.965 output. The USD numeric fields stay null because we do not convert a non-USD first-party price.

## Workload-aware monthly cost example

10 conversations per day × 8 messages × 22 active days, 1200 input and 400 output tokens per message, no cache. Derived here from the published rates above by this site's own calculator — not a figure any source published.

| Field | Value |
| --- | --- |
| Modelled monthly cost | Unavailable |
| Modelled tokens | Unavailable |
| Reason | The applicable input rate is unavailable. |

## Benchmark record

84 matched benchmark rows with their published value, unit, and provenance.

### Knowledge

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| GPQA (Graduate-Level Google-Proof Q&A) | 92.6 | 448 questions | Graduate level | [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://arxiv.org/abs/2311.12022) |
| GPQA-D (GPQA Diamond) | 92.6 | Graduate-level science questions | Graduate level | [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://arxiv.org/abs/2311.12022) |
| HLE (Humanity's Last Exam) | 43.6 | Expert-level questions | Frontier expert level | [Humanity's Last Exam](https://lastexam.ai/) |
| HLE w/o tools (Humanity's Last Exam without tools) | 43.6 | Expert-level questions | Frontier expert level | [Introducing GPT-5.4 mini and nano](https://openai.com/index/introducing-gpt-5-4-mini-and-nano/) |
| GPQA Diamond (Vals) (GPQA Diamond, Vals AI run) | 93.7 | Graduate-level science questions | Expert reasoning | [Vals AI GPQA Diamond, Vals AI run leaderboard](https://www.vals.ai/benchmarks/gpqa) |
| MMLU-Pro (Vals) (MMLU-Pro, Vals AI run) | 88.6 | Academic multiple-choice questions | Broad academic knowledge | [Vals AI MMLU-Pro, Vals AI run leaderboard](https://www.vals.ai/benchmarks/mmlu_pro) |

### Coding

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| Terminal-Bench 2.1 | 86.6 | Terminal-based software-agent tasks | Professional software engineering | [Terminal-Bench](https://www.tbench.ai/) |
| SWE-bench Pro | 67.7 | 1,865 repository problems | Long-horizon professional engineering | [SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?](https://arxiv.org/abs/2509.16941) |
| VulcanBench v3 | 81.2 | 23 post-cutoff repository tasks in the v3 report | Professional multi-file software engineering | [VulcanBench](https://github.com/morganlinton/VulcanBench/tree/main) |
| OpenHarmony Bench v1.0 | 60.8 | 153 app-development and bug-fix tasks | End-to-end OpenHarmony application development | [OpenHarmony Bench: Evaluating LLMs and Coding Agents on OpenHarmony App Development](https://arxiv.org/abs/2608.16022) |
| FrontierSWE | 73.5 | 17 ultra-long-horizon engineering and research tasks | Ultra-long-horizon frontier software engineering | [FrontierSWE: Benchmarking coding agents at the limits of human abilities](https://www.frontierswe.com/blog) |
| FrontierSWE v2 | 15.8 | 34 ultra-long-horizon engineering and research tasks | Ultra-long-horizon frontier software engineering | [FrontierSWE v2](https://www.frontierswe.com/blog/v2) |
| Bug Hunt Bench | 25.7 | 105 planted bugs across two production TypeScript repositories | Blind production-repository bug finding and repair | [Bug Hunt Bench method, caveats, and definitions](https://bughunt.productcompass.pm/method) |
| MLS-Bench Lite | 41.0 | 30 machine-learning research tasks | ML research and systems engineering | [MLS-Bench](https://mls-bench.com/) |
| PaperBench | 93.0 | AI research-paper reproduction | Frontier autonomous research and engineering | [Qwen3.8-Max: A New Bar for Coding and Cowork](https://qwen.ai/blog?id=qwen3.8) |
| QwenReactBench | 1724 | Bilingual React project construction | Production frontend development | [Qwen3.8-Max: A New Bar for Coding and Cowork](https://qwen.ai/blog?id=qwen3.8) |
| NL2Repo | 55.9 | Natural language to repository tasks | System-level software comprehension | [MiniMax M2.7: Early Echoes of Self-Evolution](https://www.minimax.io/news/minimax-m27-en) |
| LiveCodeBench (Vals) (LiveCodeBench, Vals AI run) | 87.9 | Competitive programming problems (easy, medium, hard) | Frontier coding | [Vals AI LiveCodeBench, Vals AI run leaderboard](https://www.vals.ai/benchmarks/lcb) |
| SWE-bench (Vals) (SWE-bench, Vals AI run) | 85.6 | Real repository issues by human time bucket | Frontier coding agents | [Vals AI SWE-bench, Vals AI run leaderboard](https://www.vals.ai/benchmarks/swebench) |
| DeepSWE | 56.6 | 113 software engineering tasks across 91 repositories and 5 languages | Long-horizon software engineering | [DeepSWE benchmark blog](https://deepswe.datacurve.ai/blog) |
| OpenHarmony Bench v1.2 | 67.77 | App-development, bug-fix and migration tasks (count not published) | End-to-end OpenHarmony application development and migration | [OpenHarmony Bench: Evaluating LLMs and Coding Agents on OpenHarmony App Development](https://arxiv.org/abs/2608.16022) |
| ScarfBench (ScarfBench: Cross-Framework Application Migration in Enterprise Java) | 63.7 | 204 framework migration tasks across 34 enterprise Java applications | Whole-application framework migration | [ScarfBench: A Benchmark for Cross-Framework Application Migration in Enterprise Java](https://arxiv.org/abs/2605.06754) |

### Mathematics

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| FrontierMath v2 (Tiers 1-3) (FrontierMath v2 Tiers 1-3) | 74.737 | 285 private advanced mathematics problems | From olympiad-plus to early research mathematics | [FrontierMath v2 benchmark hub](https://epoch.ai/benchmarks/frontiermath-tiers-1-3-v2) |
| FrontierMath v2 (Tier 4) (FrontierMath v2 Tier 4) | 46.341 | 41 private extreme-difficulty mathematics problems | Research-level mathematics requiring hours or days of expert work | [FrontierMath Tier 4 v2 leaderboard](https://epoch.ai/benchmarks/frontiermath-tier-4-v2?view=graph&tab=leaderboard) |

### Reasoning

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| LongBench v2 | 66.3 | Long-context tasks | Hard long-context | [LongBench v2](https://arxiv.org/abs/2412.15204) |
| MRCRv2 | 92.9 | Long-context retrieval | Hard long-context | [Introducing GPT-5.2 and GPT-5.2 Pro](https://openai.com/index/introducing-gpt-5-2/) |

### Instruction Following

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| IFBench (Instruction Following Benchmark) | 82.8 | — | — | [BenchLM](https://benchlm.ai/benchmarks/ifbench) |

### Agentic

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| Terminal-Bench 2.1 | 86.6 | Terminal-based software-agent tasks | Professional software engineering | [Terminal-Bench](https://www.tbench.ai/) |
| HLE w/ tools (Humanity's Last Exam with tools) | 56.2 | Expert questions with tool use | Frontier tool-augmented reasoning | [DeepSeek-V4 Technical Report](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf) |
| OSWorld-Verified | 86.1 | 369 real-world computer tasks (361 when eight Google Drive tasks are excluded) | Multi-step desktop and cross-application workflows | [OSWorld](https://os-world.github.io/) |
| OSWorld 2.0 | 19.4 | 108 long-horizon computer-use workflows | Long-horizon professional workflows | [OSWorld2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks](https://arxiv.org/abs/2606.29537) |
| JobBench | 53.4 | 130 tasks across 35 occupations | Professional multi-source workflows | [JobBench: Aligning Agent Work With Human Will](https://arxiv.org/abs/2605.26329) |
| AndroidWorld | 85.3 | Android app workflows | Complex mobile task completion | [GLM-5V-Turbo](https://docs.z.ai/guides/vlm/glm-5v-turbo) |
| Toolathlon-Verified | 72.5 | Verified multi-tool workflows | Advanced tool use | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| AutomationBench | 27.3 | 600 public automation tasks | Long-horizon automation | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| Agents' Last Exam | 52.4 | Agent tasks | Advanced agentic work | [DeepSeek V4 Flash 0731 update](https://api-docs.deepseek.com/zh-cn/updates/) |
| WideResearch | 81.9 | Open-ended research tasks | Broad research-agent workflows | [Qwen3.6 launch benchmarks](https://qwen.ai/blog?id=qwen3.6) |
| CoWorkBench | 74.8 | Long-horizon professional workflows | Cross-domain professional work | [Qwen3.8-Max: A New Bar for Coding and Cowork](https://qwen.ai/blog?id=qwen3.8) |
| MobileWorld | 77.8 | Interactive mobile-device workflows | Long-horizon mobile computer use | [Qwen3.8-Max: A New Bar for Coding and Cowork](https://qwen.ai/blog?id=qwen3.8) |
| WebArena-Verified (WebArena-Verified Browser Agent Benchmark) | 66.8 | 812 verified tasks; separate 258-task Hard subset | Audited stateful browser work | [WebArena-Verified: A Fully Audited Benchmark for Web Agents](https://openreview.net/forum?id=94tlGxmqkN) |
| Terminal-Bench 2.1 (Vals) (Terminal-Bench 2.1, Vals AI run) | 67.4 | Difficult terminal tasks | Frontier agentic | [Vals AI Terminal-Bench 2.1, Vals AI run leaderboard](https://www.vals.ai/benchmarks/terminal-bench-2-1) |
| Hermes Index (Hermes Index: agentic capability and cost per task inside Hermes Agent) | 36.2 | Four suites: Hermes Bench (150 tasks), TerminalBench 4 without its 4 GPU tasks, TerminalBench Science and SkillsBench | Frontier agentic | [Hermes Index: an agentic model leaderboard](https://portal.nousresearch.com/bench) |
| Hermes Bench (Hermes Bench: 150 agent tasks graded on the files and state the agent leaves behind) | 70.5 | 150 tasks across 25 categories | Frontier agentic | [Hermes Bench (Hermes Index)](https://portal.nousresearch.com/bench) |
| MazeBench (MazeBench: visual spatial reasoning in a 3D open world) | 0 | 100 hidden gems across 256 rooms in one open world | Long-horizon spatial planning, with single puzzles that can take over one hundred moves | [MazeBench Results](https://mazebench.com/blog?post=maze-bench-results) |
| CWE-bench (CWE-bench v2) | 45.4 | Audit and patch C/C++, Go, Java, TypeScript/JavaScript, Python, and Rust repositories; 100 public and 150 held-out tasks | Frontier agent evaluation | [CWE-bench v2 leaderboard and methodology](https://cwe-bench.com/) |

### Multimodal & Grounded

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| MMMU-Pro (Massive Multi-discipline Multimodal Understanding Pro) | 82.3 | Multimodal academic reasoning | Frontier multimodal | [MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark](https://arxiv.org/abs/2409.02813) |
| OCRBench V2 | 74.2 | Image OCR tasks | Native visual text understanding | [OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning](https://arxiv.org/abs/2501.00321) |
| MathVision w/ Python (MathVision with Python) | 97.7 | Visual mathematics problems with Python | Advanced multimodal mathematics | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| BabyVision w/ Python (BabyVision with Python) | 91.3 | Visual perception tasks with Python | Fine-grained visual perception | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| ZeroBench w/ Python (ZeroBench_main with Python) | 49.0 | Visual reasoning questions with Python | Tool-augmented visual reasoning | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| PerceptionBench (PerceptionBench (Internal)) | 63.5 | Internal atomic visual-perception tasks | Fine-grained visual perception | [Kimi K3: Open Frontier Intelligence](https://www.kimi.com/blog/kimi-k3) |
| OmniDocBench 1.5 | 92.1 | Document understanding tasks | Grounded document reasoning | [Introducing GPT-5.4 mini and nano](https://openai.com/index/introducing-gpt-5-4-mini-and-nano/) |
| RealWorldQA | 88.0 | Real-world visual question answering | General visual reasoning | [Qwen3.6 launch benchmarks](https://qwen.ai/blog?id=qwen3.6) |
| Video-MME (with subtitle) (Video-MME with subtitle) | 90.4 | Video understanding | Multimodal video reasoning | [Qwen3.6 launch benchmarks](https://qwen.ai/blog?id=qwen3.6) |
| MathVision | 95.2 | Visually grounded math problems | Advanced multimodal mathematics | [Qwen3.6 launch benchmarks](https://qwen.ai/blog?id=qwen3.6) |
| CC-OCR | 79.6 | Optical character recognition | Document reading | [Qwen3.6 launch benchmarks](https://qwen.ai/blog?id=qwen3.6) |
| ERQA | 77.8 | Evidence-based visual QA | Grounded multimodal reasoning | [Qwen3.6 launch benchmarks](https://qwen.ai/blog?id=qwen3.6) |
| VideoMMMU | 88.7 | Video-grounded expert reasoning | Frontier multimodal video reasoning | [Qwen3.6 launch benchmarks](https://qwen.ai/blog?id=qwen3.6) |
| MLVU (M-Avg) (MLVU mean average) | 90.8 | General video understanding | Broad multimodal video reasoning | [Qwen3.6 launch benchmarks](https://qwen.ai/blog?id=qwen3.6) |
| LVBench | 81.8 | Long-form video question answering | Extended temporal reasoning | [Qwen3.8-Max: A New Bar for Coding and Cowork](https://qwen.ai/blog?id=qwen3.8) |
| MMVU (Multimodal Multi-disciplinary Video Understanding) | 82.4 | Video understanding | Multi-disciplinary multimodal video reasoning | [Kimi K2.5 benchmark release surface](https://www.kimi.com/blog/kimi-k2-5.html) |
| ScreenSpot Pro | 84.5 | 1,581 grounding instructions | Professional GUI grounding | [ScreenSpot-Pro: GUI Grounding for Professional High-Resolution Computer Use](https://arxiv.org/abs/2504.07981) |
| MedXpertQA (MM) (MedXpertQA Multimodal) | 80.4 | 2,000 multimodal medical questions | Clinical multimodal reasoning | [Muse Spark Eval Methodology](https://ai.meta.com/static-resource/muse-spark-eval-methodology) |
| ZeroBench | 24.0 | 100 visual reasoning questions | Tool-augmented visual reasoning | [Muse Spark Eval Methodology](https://ai.meta.com/static-resource/muse-spark-eval-methodology) |
| Vision2Web | 69.0 | Screenshot-to-web tasks | Multimodal web generation | [GLM-5V-Turbo](https://docs.z.ai/guides/vlm/glm-5v-turbo) |
| SimpleVQA | 75.0 | Visual QA tasks | General visual understanding | [GLM-5V-Turbo](https://docs.z.ai/guides/vlm/glm-5v-turbo) |
| CharXiv (CharXiv Reasoning) | 93.5 | Scientific chart reasoning | Scientific visualization reasoning | [CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs](https://charxiv.github.io/) |
| CharXiv w/o tools (CharXiv Reasoning without tools) | 88.4 | Scientific chart reasoning (tool-free) | Scientific visualization reasoning | [CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs](https://charxiv.github.io/) |
| BabyVision | 82.0 | Visual perception tasks | Fine-grained visual perception | [Muse Spark 1.1 Evaluation Report](https://ai.meta.com/static-resource/muse-spark-1-1-evaluation-report) |

### external

| Benchmark | Value | Tasks | Difficulty | Provenance |
| --- | --- | --- | --- | --- |
| Chess Puzzles (Chess Puzzles (Epoch AI)) | 46.0 | 100 programmatically generated chess positions with a single best move | Novel chess positions, single best move | [Chess Puzzles (Epoch AI)](https://epoch.ai/benchmarks/chess-puzzles) |
| Mystery Game Puzzles (Mystery Game Puzzles (Epoch AI)) | 38.0 | 100 single-best-move positions from an undisclosed game | Unfamiliar game positions, single best move | [Mystery Game Puzzles (Epoch AI)](https://epoch.ai/benchmarks/mystery-game-puzzles) |
| MirrorCode (MirrorCode (Epoch AI and METR)) | 52.2 | 30 reimplementation tasks (15 programs in two languages), three attempts each | Days-long software engineering without source code | [MirrorCode (Epoch AI)](https://epoch.ai/benchmarks/mirrorcode) |
| Epoch Automation Reports | 53.0 | 11 open-ended work tasks in five categories from Epoch AI's own work | Open-ended professional research and design work | [Epoch Automation Reports](https://epoch.ai/benchmarks/epoch-automation-reports) |
| FACTS Multimodal | 51.5 | About 1,500 image questions: 711 public and 811 held out | Image-grounded factual answering | [BenchLM](https://benchlm.ai/benchmarks/facts-multimodal) |
| Code2Video | 1533 | Creative briefs rendered as motion-graphics video | Code-driven motion design | [BenchLM](https://benchlm.ai/benchmarks/code2video) |
| Landing Page Arena | 1587 | 12 client briefs for small-business landing pages, rated by human votes | Marketing page design and copy | [BenchLM](https://benchlm.ai/benchmarks/landing-page-arena) |
| Joke Arena | 1456 | 100 joke prompts in 10 categories, rated by human votes | Humour writing | [BenchLM](https://benchlm.ai/benchmarks/joke-arena) |
| Pelican Arena | 1583 | One prompt: an animated SVG of a pelican riding a bicycle, rated by human votes | Vector drawing and animation in code | [BenchLM](https://benchlm.ai/benchmarks/pelican-arena) |
| Landmark Arena | 1474 | 20 famous landmarks built as three.js web pages, rated by human votes | 3D scenes in browser code | [BenchLM](https://benchlm.ai/benchmarks/landmark-arena) |
| ASCII Arena | 1262 | ASCII drawings of subjects from Google's Quick, Draw! dataset, rated by human votes | Text-character drawing | [BenchLM](https://benchlm.ai/benchmarks/ascii-arena) |
| BabaIsBench (BabaIsBench (Baba Is You intro levels, pass@1)) | 58 | 8 intro levels of Baba Is You, three attempts per level (Stage 0); 15 further levels run once (Stage 1) | Rule-rewriting puzzles that need planning over many moves | [Baba Is Solved by Fable 5 and GPT-5.6 Sol, but at what cost? (Quesma)](https://quesma.com/blog/baba-is-bench/) |
| NYT Connections (NYT Connections LLM Benchmark (extended version)) | 88.3 | 940 NYT Connections puzzles, each with up to four added trick words | Wordplay, word knowledge and decoys that fit almost-but-not-quite | [NYT Connections LLM Benchmark (Lech Mazur, GitHub)](https://github.com/lechmazur/nyt-connections) |
| StealthMark (StealthMark Intelligence Benchmarks (overall score)) | 62.4 | 8 original tasks: Newton's Mirror, Ground Truth, Latch, Foresight, Mona Lisa, Selfie, Beach and Raccoon | Chained physical reasoning, planning without tools and precise visual construction | [StealthMark: The AI Benchmark Explained (Stealth Models)](https://stealthmodels.com/stealthmark) |
| HoneyBench (HoneyBench v0.2 (Goodhart Labs)) | 66.2 | 20 private honeypot tasks across 7 domains with exploitable evaluators | Agentic alignment and specification gaming | [HoneyBench v0.2](https://goodhartlabs.com/honeybench) |

## Independent scores (Epoch AI)

A second, independent score source: Epoch's own figures in Epoch's units, one line per model version Epoch evaluated (effort setting as Epoch labels it); higher is better. Not part of this site's capability scores or rank.

| Benchmark | Variant | Score | Unit | Epoch date | Evidence |
| --- | --- | --- | --- | --- | --- |
| GPQA Diamond | xhigh | 0.9268 | fraction (0–1) | 2026-08-04 | Observed 2026-11-21 · source epoch:benchmarks |
| FrontierMath Tiers 1-3 (v2) | xhigh | 0.7474 | fraction (0–1) | 2026-08-04 | Observed 2026-11-21 · source epoch:benchmarks |
| FrontierMath Tier 4 (v2) | xhigh | 0.4634 | fraction (0–1) | 2026-08-04 | Observed 2026-11-21 · source epoch:benchmarks |
| SimpleQA Verified | xhigh | 0.458 | fraction (0–1) | 2026-08-27 | Observed 2026-11-21 · source epoch:benchmarks |

Source: [Epoch AI, CC BY 4.0](https://epoch.ai/benchmarks) ([licence](https://creativecommons.org/licenses/by/4.0/)). Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from https://epoch.ai/benchmarks.

## Lifecycle and limitations log

Lifecycle events the source associates with this model.

No lifecycle event references this model. That is not evidence the model has no lifecycle plan — only that this source published none.

## What this profile does not claim

Values are reproduced exactly as their sources published them, in the units those sources declared; none are converted, interpolated, or averaged across providers. Any field marked unavailable was attempted and not returned. Last attempted fetch for this model's score: 2026-11-21 05:17 UTC.
