Benchmark leaderboard - 20 entries
GPQA Diamond leaderboard
GPQA is a dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry, and GPQA Diamond is its 198-question subset where both expert validators answered correctly and at most one of three non-experts did. As of November 7, 2026, GPT-6 Astra is listed first on GPQA Diamond with a vendor-reported 96.0%. Every number here is reported by the model's own vendor under its own settings, so the order is not a controlled comparison.
Last verified from 8 sources. Read the caveats
GPQA Diamond
20 entries, in the board's order
| 1 | GPT-6 Astra OpenAI research environment or API; 'Evaluation scores are the maximum at any effort.' Vendor-reported | 96.0% |
| 2 | GPT-5.6 Sol OpenAI GPT-5.6 launch eval table (same value repeated in the GPT-6 Astra table) Vendor-reported | 94.6% |
| 3 | Claude Mythos Preview Anthropic system card: 'averaged over 5 trials' Vendor-reported | 94.55% |
| 4 | Gemini 3.1 Pro Gemini 3.1 Pro Thinking (High), No tools Vendor-reported | 94.3% |
| 5 | Claude Opus 4.7 Anthropic system card: 'averaged over 10 trials' Vendor-reported | 94.2% |
| 6 | Claude Mythos 5 Claude Fable 5 & Claude Mythos 5 System Card section 8.8: 'averaged over 5 trials'; Anthropic notes it considers GPQA Diamond saturated and plans to stop reporting it (same 94.1% shown for Claude Mythos 5 in OpenAI's GPT-5.6 table) Vendor-reported | 94.1% |
| 7 | GPT-5.5 'Evals of GPT were run with reasoning effort set to xhigh' in a research environment Vendor-reported | 93.6% |
| 8 | Claude Opus 4.8 Anthropic system card: 'averaged over 25 trials' Vendor-reported | 93.6% |
| 9 | GPT-5.6 Terra OpenAI GPT-5.6 launch eval table Vendor-reported | 92.9% |
| 10 | Qwen3.8-Max Qwen3.8-Max Full Benchmark Table; the post says 2.4T parameters (95B active), with open weights releasing next week Vendor-reported | 92.6 |
| 11 | DeepSeek-V4-Pro GPQA Diamond (Pass@1), 'Comparison with frontier models (Max reasoning effort)' table in the DeepSeek-V4.1-Flash model card Vendor-reported | 92.4 |
| 12 | GPT-5.6 Luna OpenAI GPT-5.6 launch eval table Vendor-reported | 92.3% |
| 13 | Gemini 3 Pro Gemini 3 launch post: 'GPQA Diamond (91.9%)'; the Gemini 3.1 Pro model card lists Gemini 3 Pro Thinking (High), No tools at 91.9% Vendor-reported | 91.9% |
| 14 | Qwen3.8-Flash-Next Qwen3.8-Flash-Next blog Performance table (Scientific reasoning) Vendor-reported | 91.7 |
| 15 | GLM-5.2 Z.ai Full Benchmark Table; reasoning tasks use temperature=1.0, top_p=0.95, max generation length 163,840 tokens; MIT license, weights on Hugging Face Vendor-reported | 91.2 |
| 16 | DeepSeek-V4.1-Flash GPQA Diamond (Pass@1), 'Comparison with frontier models (Max reasoning effort)' table; instruct results use reasoning_effort=100, temperature=1.0, top_p=0.95; MIT-licensed weights Vendor-reported | 90.9 |
| 17 | Gemini 3 Flash Gemini 3 Flash launch post Vendor-reported | 90.4% |
| 18 | Qwen3.8-27B Qwen3.8-27B model card benchmark table (Scientific reasoning) Vendor-reported | 89.2 |
| 19 | Muse Spark Open (816B-A42B) Muse Spark Open (xhigh): 'xhigh reasoning effort, temperature = 1.0, top_p=1.0, top_k=disabled across benchmarks'; open weights Vendor-reported | 89.0 |
| 20 | Gemma 4 31B Gemma 4 model card benchmark results (instruction-tuned models) Vendor-reported | 84.3% |
About this board
- Rows marked vendor-reported come from the model maker's own model card or launch post, run under its own settings. Read the setting under each model before comparing two rows.
- All scores are vendor-reported from model cards, system cards and release posts; third-party aggregator leaderboards were excluded. The benchmark paper is arXiv:2311.12022.
- The Claude Fable 5 & Claude Mythos 5 System Card (June 9, 2026) says Anthropic considers GPQA Diamond a saturated evaluation and plans to stop reporting the performance of future models on it.
- OpenAI's GPT-6 Astra table lists Claude Fable 5.1 at 93.7%, Claude Fable 5 at 92.6%, Claude Opus 5 at 93.7% and Gemini 3.8 Flash at 95.3%. Those are OpenAI's figures, not the model makers' own, so they are not listed here.
- GLM-5.3: Meta's Muse Spark Open card lists GLM 5.3 (max) at 91.7 and DeepSeek's V4.1-Flash card lists GLM-5.3 at 88.1. Neither is Z.ai's own figure, so neither is listed here.
- MiniMax M3: Z.ai's GLM-5.2 table lists it at 93.0. That is Z.ai's figure, not MiniMax's, so it is not listed here.
- No release date was found on the pages read for GPT-6 Astra (the post carries update notes dated September 22 and 29, 2026), Claude Opus 4.7, DeepSeek-V4.1-Flash and DeepSeek-V4-Pro, Qwen3.8-27B (its citation gives the month only) and the Gemma 4 model card (last updated 2026-07-30), so those dates are left blank.
FAQ
What does GPQA Diamond measure?
GPQA is a dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry, and GPQA Diamond is its 198-question subset where both expert validators answered correctly and at most one of three non-experts did.
Which model leads GPQA Diamond?
As of November 7, 2026, GPT-6 Astra is listed first on GPQA Diamond with a vendor-reported 96.0%. Vendors run the test under their own settings, so the list order is not a controlled comparison.
Where do these GPQA Diamond scores come from?
20 are vendor-reported, taken from the model maker's own model card or launch post. Every entry links to its source, which was read on November 7, 2026, and lists the setting it was run under.
Related reading
- DeepInfer: What DeepSeek Open-Sourced and What You Can RunDeepSeek open-sourced DeepInfer, the inference engine behind its API. It is offline-only, V4.1-Flash-only and Blackwell-or-Ascend-only. What it is, what is missing, and who benefits.
- LFM2.5-24B-A2B vs Qwen3.6-35B-A3B: Fast Local Agent or Better Coder?LFM2.5-24B-A2B vs Qwen3.6-35B-A3B for local agents: Liquid wins tool-use benchmarks, Qwen wins SWE-Bench Verified. Which one to run for which job.
- Claude Advisor Tool vs a Bigger Model: When the Pairing PaysAnthropic's own runs: a Fable 5.1 advisor added 1.7 points to Opus 5.5 for 2.1x the cost. When the Claude advisor tool pays, and when to raise effort instead.
- OpenAI Decisions API vs Jev vs Clef: Price, Speed, Which to PickOpenAI Decisions API costs $0.10 per million input tokens, 2.4x Jev. Pick Jev for cheap text decisions, Clef-flash for speed, Decisions API for images.
- Run Qwen3.8-Flash-Next on a Gaming PC With StrataRun Qwen3.8-Flash-Next locally with Strata on a 12 GB GPU: hardware needs, speeds by quant, the license catch, and when the hosted API is the better call.
- Kolibri Release Guide: Aleph Alpha's 78B Open ModelKolibri is an Apache 2.0 English-German MoE with 78B total and 3B active parameters and a 1M context. Benchmarks, the vLLM command, and who should run it.
Sources
Last verified . Each row links to the page its score was copied from.
- https://openai.com/index/gpt-6-astra/
- https://openai.com/index/gpt-5-6/
- https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf
- https://deepmind.google/models/model-cards/gemini-3-1-pro/
- https://www-cdn.anthropic.com/037f06850df7fbe871e206dad004c3db5fd50340/Claude%20Opus%204.7%20System%20Card.pdf
- https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620.pdf
- https://openai.com/index/introducing-gpt-5-5/
- https://www-cdn.anthropic.com/0b4915911bb0d19eca5b5ee635c80fef830a37ea.pdf
- https://qwen.ai/blog?id=qwen3.8
- https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- https://blog.google/products-and-platforms/products/gemini/gemini-3/
- https://qwen.ai/blog?id=qwen3.8-flash-next
- https://z.ai/blog/glm-5.2
- https://blog.google/products-and-platforms/products/gemini/gemini-3-flash/
- https://huggingface.co/Qwen/Qwen3.8-27B
- https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B
- https://ai.google.dev/gemma/docs/core/model_card_4
Newsletter
Get new benchmark results as they land
One email per week. New model releases, the leaderboard moves that matter and what they mean for what you build.