| group,benchmark,subtitle,torki_2,torki_setting,model,lab,score,source_type,conditions,conditions_short,source_url,source_date
|
| Anthropic and OpenAI,GPQA Diamond,Graduate-level science,89.65,,Claude Opus 5.5,Anthropic,90.6,third-party: Epoch AI,"Epoch AI run ""claude-opus-5-5_max"" (effort/budget: max); 198 questions, simple-evals prompt, strict ""ANSWER: LETTER"" parsing (unparsed = wrong), Inspect harness; stderr ±1.94 pp.",max,https://epoch.ai/benchmarks/gpqa-diamond,2026-09-22
|
| Anthropic and OpenAI,GPQA Diamond,Graduate-level science,89.65,,Claude Fable 5.1,Anthropic,93.7,third-party: Artificial Analysis,"AA variant ""Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback)"" (effort: max); 198 questions, 4-option MC, regex extraction, pass@1 (repeats not stated for this legacy eval); temperature 0.6 for reasoning models unless lab recommends otherwise; max output tokens = lab maximum.",max,https://artificialanalysis.ai/models/claude-fable-5-1,unknown (no date on page; fetched 2026-10-01)
|
| Anthropic and OpenAI,GPQA Diamond,Graduate-level science,89.65,,Claude Sonnet 5.5,Anthropic,95.6,third-party: Epoch AI,"Epoch AI run ""claude-sonnet-5-5_max"" (effort/budget: max); 198 questions, simple-evals prompt, strict ""ANSWER: LETTER"" parsing (unparsed = wrong), Inspect harness; stderr ±1.37 pp.",max,https://epoch.ai/benchmarks/gpqa-diamond,2026-09-29
|
| Anthropic and OpenAI,GPQA Diamond,Graduate-level science,89.65,,Claude Haiku 4.5,Anthropic,73.0,official,"Extended thinking with 128K thinking budget, default sampling (temperature, top_p). 'All other scores were averaged over 10 runs'.",,https://www.anthropic.com/news/claude-haiku-4-5,2025-10-15
|
| Anthropic and OpenAI,GPQA Diamond,Graduate-level science,89.65,,GPT-6 Astra,OpenAI,96.0,official,Headline table value; footnote: 'Evaluation scores are the maximum at any effort.' Max effort per the page's chart data (0.959596).,max,https://openai.com/index/gpt-6-astra/,2026-09-03
|
| Anthropic and OpenAI,GPQA Diamond,Graduate-level science,89.65,,GPT-6.1 Sol,OpenAI,95.4,third-party: Epoch AI,"Epoch AI run ""gpt-6.1-sol_max"" (effort/budget: max); 198 questions, simple-evals prompt, strict ""ANSWER: LETTER"" parsing (unparsed = wrong), Inspect harness; stderr ±1.38 pp.",max,https://epoch.ai/benchmarks/gpqa-diamond,2026-09-29
|
| Anthropic and OpenAI,GPQA Diamond,Graduate-level science,89.65,,GPT-6 Luna,OpenAI,90.5,third-party: Epoch AI,"Epoch AI run ""gpt-6-luna_max"" (effort/budget: max); 198 questions, simple-evals prompt, strict ""ANSWER: LETTER"" parsing (unparsed = wrong), Inspect harness; stderr ±1.85 pp.",max,https://epoch.ai/benchmarks/gpqa-diamond,2026-09-22
|
| Anthropic and OpenAI,GPQA Diamond,Graduate-level science,89.65,,gpt-oss-120b,OpenAI,80.1,official,Reasoning effort high; 'basic pass@1 results ... using the model's default system prompt'.,high,https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf,2025-08-05
|
| Anthropic and OpenAI,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Claude Opus 5.5,Anthropic,61.4,third-party: Artificial Analysis,"Artificial Analysis run, the ""Max Effort"" variant (max reasoning effort): 0.6135; 2,158 text-only questions (multimodal excluded), pass@1, LLM equality checker; no tools. Re-verified 2026-10-02.",max,https://artificialanalysis.ai/evaluations/humanitys-last-exam,unknown
|
| Anthropic and OpenAI,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Claude Fable 5.1,Anthropic,59.1,third-party: Artificial Analysis,"Artificial Analysis run, the ""Max Effort"" variant (max reasoning effort): 0.5913; 2,158 text-only questions (multimodal excluded), pass@1, LLM equality checker; no tools. Re-verified 2026-10-02.",max,https://artificialanalysis.ai/evaluations/humanitys-last-exam,unknown
|
| Anthropic and OpenAI,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Claude Sonnet 5.5,Anthropic,55.0,third-party: Artificial Analysis,"AA variant ""Claude Sonnet 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)"" (effort: max); 2,158 text-only questions of the 2,500 (May 2025 revision), 1 repeat, pass@1, GPT-5.6 Luna (medium) equality checker; temperature 0.6 for reasoning models unless lab recommends otherwise; max output tokens = lab maximum.",max,https://artificialanalysis.ai/models/claude-sonnet-5-5,unknown (no date on page; fetched 2026-10-01)
|
| Anthropic and OpenAI,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Claude Haiku 4.5,Anthropic,10.4,third-party: Artificial Analysis,"AA variant ""Claude 4.5 Haiku (Reasoning)"" (effort: n/a); 2,158 text-only questions of the 2,500 (May 2025 revision), 1 repeat, pass@1, GPT-5.6 Luna (medium) equality checker; temperature 0.6 for reasoning models unless lab recommends otherwise; max output tokens = lab maximum.",,https://artificialanalysis.ai/models/claude-4-5-haiku-reasoning,unknown (no date on page; fetched 2026-10-01)
|
| Anthropic and OpenAI,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,GPT-6 Astra,OpenAI,54.7,third-party: Artificial Analysis,"Artificial Analysis independent run, reasoning effort max (AA variant name 'GPT-6 Astra (max)').",max,https://artificialanalysis.ai/models/gpt-6-astra,unknown
|
| Anthropic and OpenAI,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,GPT-6.1 Sol,OpenAI,52.9,third-party: Artificial Analysis,"Artificial Analysis independent run, reasoning effort max (AA variant name 'GPT-6.1 Sol (max)').",max,https://artificialanalysis.ai/models/gpt-6-1-sol,unknown
|
| Anthropic and OpenAI,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,GPT-6 Luna,OpenAI,38.5,third-party: Artificial Analysis,"Artificial Analysis independent run, reasoning effort max (AA variant name 'GPT-6 Luna (max)').",max,https://artificialanalysis.ai/models/gpt-6-luna,unknown
|
| Anthropic and OpenAI,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,gpt-oss-120b,OpenAI,19.6,third-party: Artificial Analysis,"AA variant ""gpt-oss-120b (High)"" (effort: high); 2,158 text-only questions of the 2,500 (May 2025 revision), 1 repeat, pass@1, GPT-5.6 Luna (medium) equality checker; temperature 0.6 for reasoning models unless lab recommends otherwise; max output tokens = lab maximum.",high,https://artificialanalysis.ai/models/gpt-oss-120b,unknown (no date on page; fetched 2026-10-01)
|
| Anthropic and OpenAI,MMLU-Pro,Broad expert knowledge,88.07,sample,Claude Fable 5.1,Anthropic,92.38,third-party: Vals AI,"Vals AI run; effort: max; temperature 1; max output tokens 128000; provider Anthropic; 5-shot CoT (original paper prompt), all ~12k questions; stderr ±0.266.",max,https://www.vals.ai/benchmarks/mmlu_pro,2026-09-01
|
| Anthropic and OpenAI,MMLU-Pro,Broad expert knowledge,88.07,sample,Claude Haiku 4.5,Anthropic,78.72,third-party: Vals AI,"Vals AI run; effort: thinking variant, budget not stated; temperature 1; max output tokens 64000; provider Anthropic; 5-shot CoT (original paper prompt), all ~12k questions; stderr ±0.477.",,https://www.vals.ai/benchmarks/mmlu_pro,2026-09-01
|
| Anthropic and OpenAI,MMLU-Pro,Broad expert knowledge,88.07,sample,gpt-oss-120b,OpenAI,79.17,third-party: Vals AI,"Vals AI run; effort: not stated; temperature 1; max output tokens 128000; provider Fireworks AI; 5-shot CoT (original paper prompt), all ~12k questions; stderr ±0.395.",,https://www.vals.ai/benchmarks/mmlu_pro,2026-09-01
|
| Anthropic and OpenAI,AIME 2025,"Competition maths, 2025",96.25,,Claude Haiku 4.5,Anthropic,80.7,official,"Extended thinking with 128K thinking budget, default sampling (temperature, top_p). Average over 10 independent runs that each calculate pass@1 over 16 trials. Labelled '(no tools)'.",,https://www.anthropic.com/news/claude-haiku-4-5,2025-10-15
|
| Anthropic and OpenAI,AIME 2025,"Competition maths, 2025",96.25,,gpt-oss-120b,OpenAI,92.5,official,Reasoning effort high; 'basic pass@1 results ... using the model's default system prompt'.,high,https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf,2025-08-05
|
| Anthropic and OpenAI,MATH-500,Competition maths,99.4,,gpt-oss-120b,OpenAI,94.8,third-party: Vals AI,"Vals AI run; effort: not stated; temperature 1; max output tokens 128000; provider Fireworks AI; 500 problems, boxed-answer prompt; stderr ±0.993.",,https://www.vals.ai/benchmarks/math500,2026-01-09
|
| Anthropic and OpenAI,LiveCodeBench v6,"Contest coding, 2025 problems",77.1,,Claude Fable 5.1,Anthropic,90.52,third-party: Vals AI,"Vals AI run; effort: max; temperature 1; max output tokens 128000; provider Anthropic; code generation, Python, hidden tests; overall over easy/medium/hard; stderr ±0.863.",max,https://www.vals.ai/benchmarks/lcb,2026-09-01
|
| Anthropic and OpenAI,LiveCodeBench v6,"Contest coding, 2025 problems",77.1,,Claude Haiku 4.5,Anthropic,61.5,third-party: Artificial Analysis,"AA variant ""Claude 4.5 Haiku (Reasoning)"" (effort: n/a); code_generation_lite, pass@1, no LCB custom system prompt; retired from active reporting; temperature 0.6 for reasoning models unless lab recommends otherwise; max output tokens = lab maximum.",,https://artificialanalysis.ai/models/claude-4-5-haiku-reasoning,unknown (no date on page; fetched 2026-10-01)
|
| Anthropic and OpenAI,LiveCodeBench v6,"Contest coding, 2025 problems",77.1,,gpt-oss-120b,OpenAI,83.23,third-party: Vals AI,"Vals AI run; effort: not stated; temperature 1; max output tokens 128000; provider Fireworks AI; code generation, Python, hidden tests; overall over easy/medium/hard; stderr ±1.055.",,https://www.vals.ai/benchmarks/lcb,2026-09-01
|
| Anthropic and OpenAI,IFBench,Unseen instructions,73.33,,Claude Haiku 4.5,Anthropic,54.3,third-party: Artificial Analysis,"AA variant ""Claude 4.5 Haiku (Reasoning)"" (effort: n/a); 294 single-turn questions, 5 repeats, loose mode, prompt-level accuracy, official allenai code; temperature 0.6 for reasoning models unless lab recommends otherwise; max output tokens = lab maximum.",,https://artificialanalysis.ai/models/claude-4-5-haiku-reasoning,unknown (no date on page; fetched 2026-10-01)
|
| Anthropic and OpenAI,IFBench,Unseen instructions,73.33,,gpt-oss-120b,OpenAI,69.0,third-party: Artificial Analysis,"Artificial Analysis independent run, variant '(high)' reasoning.",high,https://artificialanalysis.ai/models/gpt-oss-120b,unknown
|
| Anthropic and OpenAI,MMMU,College questions about images,84.67,,Claude Haiku 4.5,Anthropic,73.2,official,"Extended thinking with 128K thinking budget, default sampling (temperature, top_p). Averaged over 10 runs.",,https://www.anthropic.com/news/claude-haiku-4-5,2025-10-15
|
| Anthropic and OpenAI,MILU,11 Indian languages,85.09,sample,Claude Opus 5.5,Anthropic,93.1,official,Average accuracy across 11 languages (10 Indic + English); adaptive thinking at max effort; averaged over 5 trials; no tools or customized system prompts.,max,https://www.anthropic.com/claude-opus-5-5-system-card,2026-09-22
|
| Anthropic and OpenAI,MILU,11 Indian languages,85.09,sample,Claude Fable 5.1,Anthropic,93.0,official,Average accuracy across 11 languages (10 Indic + English). Adaptive thinking and max effort; averaged over five trials; no tools or customized system prompts.,max,https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system-card,2026-09-01
|
| Anthropic and OpenAI,MILU,11 Indian languages,85.09,sample,Claude Sonnet 5.5,Anthropic,91.6,official,Average accuracy across 11 languages (10 Indic + English); adaptive thinking at max effort; averaged over 5 trials; no tools or customized system prompts.,max,https://www.anthropic.com/claude-sonnet-5-5-system-card,2026-09-28
|
| Anthropic and OpenAI,MGSM,Maths in 11 languages,85.96,see note,Claude Haiku 4.5,Anthropic,92.15,third-party: Vals AI,"Vals AI run; effort: thinking variant, budget not stated; temperature 1; max output tokens 64000; provider Anthropic; simple-evals grading scripts, 11 languages (incl. bn, te), translated prompt, default temperature; stderr ±1.649.",,https://www.vals.ai/benchmarks/mgsm,2026-01-09
|
| Anthropic and OpenAI,MGSM,Maths in 11 languages,85.96,see note,gpt-oss-120b,OpenAI,92.04,third-party: Vals AI,"Vals AI run; effort: not stated; temperature 1; max output tokens 128000; provider Fireworks AI; simple-evals grading scripts, 11 languages (incl. bn, te), translated prompt, default temperature; stderr ±1.642.",,https://www.vals.ai/benchmarks/mgsm,2026-01-09
|
| "Google, Meta and xAI",GPQA Diamond,Graduate-level science,89.65,,Gemini 3.1 Pro,Google,94.3,official,Thinking (High); model-id gemini-3.1-pro-preview; default API sampling; results as of Feb 2026; 'No tools'; self computed,high,https://deepmind.google/models/gemini/pro/,2026-02
|
| "Google, Meta and xAI",GPQA Diamond,Graduate-level science,89.65,,Gemini 3 Flash,Google,90.4,official,Thinking; model-id gemini-3-flash-preview; default API sampling; benchmarks as of Dec 2025; 'No tools',,https://storage.googleapis.com/deepmind-media/gemini/gemini_3_flash_model_evaluation.pdf,2025-12
|
| "Google, Meta and xAI",AIME 2025,"Competition maths, 2025",96.25,,Gemini 3 Flash,Google,95.2,official,Thinking; model-id gemini-3-flash-preview; default API sampling; benchmarks as of Dec 2025; 'No tools',,https://storage.googleapis.com/deepmind-media/gemini/gemini_3_flash_model_evaluation.pdf,2025-12
|
| "Google, Meta and xAI",GPQA Diamond,Graduate-level science,89.65,,Gemini 3.1 Flash-Lite,Google,86.9,official,High thinking; model-id gemini-3.1-flash-lite-preview (now GA as gemini-3.1-flash-lite); default sampling; results as of Mar 2026; 'No tools'; self-computed,high,https://storage.googleapis.com/deepmind-media/gemini/gemini_3-1_flash-lite_model_evaluation.pdf,2026-03
|
| "Google, Meta and xAI",LiveCodeBench v6,"Contest coding, 2025 problems",77.1,,Gemini 3.1 Flash-Lite,Google,72.0,official,High thinking; model-id gemini-3.1-flash-lite-preview (now GA as gemini-3.1-flash-lite); default sampling; results as of Mar 2026; 'LiveCodeBench Code generation results use the 175 UI problems from the date range: 1/1/2025-5/1/2025. Results for all models are self-computed.',high,https://storage.googleapis.com/deepmind-media/gemini/gemini_3-1_flash-lite_model_evaluation.pdf,2026-03
|
| "Google, Meta and xAI",GPQA Diamond,Graduate-level science,89.65,,Gemma 4 31B,Google,84.3,official,"Instruction-tuned; thinking mode presumably ON (the Gemma 3 27B comparison column is explicitly labelled '(no think)', Gemma 4 columns are not); sampling recommended temperature=1.0",,https://ai.google.dev/gemma/docs/core/model_card_4,2026-07-30
|
| "Google, Meta and xAI",MMLU-Pro,Broad expert knowledge,88.07,sample,Gemma 4 31B,Google,85.2,official,"Instruction-tuned; thinking mode presumably ON (the Gemma 3 27B comparison column is explicitly labelled '(no think)', Gemma 4 columns are not); sampling recommended temperature=1.0",,https://ai.google.dev/gemma/docs/core/model_card_4,2026-07-30
|
| "Google, Meta and xAI",AIME 2026,"Competition maths, 2026",95.0,,Gemma 4 31B,Google,89.2,official,"Instruction-tuned; thinking mode presumably ON (the Gemma 3 27B comparison column is explicitly labelled '(no think)', Gemma 4 columns are not); sampling recommended temperature=1.0; 'no tools'",,https://ai.google.dev/gemma/docs/core/model_card_4,2026-07-30
|
| "Google, Meta and xAI",LiveCodeBench v6,"Contest coding, 2025 problems",77.1,,Gemma 4 31B,Google,80.0,official,"Instruction-tuned; thinking mode presumably ON (the Gemma 3 27B comparison column is explicitly labelled '(no think)', Gemma 4 columns are not); sampling recommended temperature=1.0; date window not stated",,https://ai.google.dev/gemma/docs/core/model_card_4,2026-07-30
|
| "Google, Meta and xAI",GPQA Diamond,Graduate-level science,89.65,,Muse Spark 1.3,Meta,93.5,third-party: Artificial Analysis,"Artificial Analysis model page, variant ""Muse Spark 1.3 (Max)"" (max reasoning effort): gpqa 0.9354; no tools. Re-verified 2026-10-02 (the launch article only gives rounded figures).",max,https://artificialanalysis.ai/models/muse-spark-1-3,2026-10-02
|
| "Google, Meta and xAI",Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Muse Spark 1.3,Meta,48.7,third-party: Artificial Analysis,"Artificial Analysis model page, variant ""Muse Spark 1.3 (Max)"" (max reasoning effort): hle 0.4870, text-only subset, no tools. Re-verified 2026-10-02 (the launch article only gives rounded figures).",max,https://artificialanalysis.ai/models/muse-spark-1-3,2026-10-02
|
| "Google, Meta and xAI",GPQA Diamond,Graduate-level science,89.65,,Muse Glimmer-30B,Meta,83.5,official,High reasoning strength; full 198 questions; card labels it 'GPQA Diamond (AA)',high,https://huggingface.co/meta-models/Muse-Glimmer-30B/raw/main/README.md,2026-08-10
|
| "Google, Meta and xAI",AIME 2026,"Competition maths, 2026",95.0,,Muse Glimmer-30B,Meta,94.7,official,High reasoning strength; full 30 AIME 2026 questions; Meta internal run,high,https://huggingface.co/meta-models/Muse-Glimmer-30B/raw/main/README.md,2026-08-10
|
| "Google, Meta and xAI",Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Muse Glimmer-30B,Meta,22.0,official,"High reasoning strength; text-only subset of 2,158 questions; no tools; card labels it 'HLE Text (AA)'",high,https://huggingface.co/meta-models/Muse-Glimmer-30B/raw/main/README.md,2026-08-10
|
| "Google, Meta and xAI",IFBench,Unseen instructions,73.33,,Muse Glimmer-30B,Meta,77.0,official,"High reasoning strength; 294 tasks, mean pass rate; Meta internal run",high,https://huggingface.co/meta-models/Muse-Glimmer-30B/raw/main/README.md,2026-08-10
|
| "Google, Meta and xAI",GPQA Diamond,Graduate-level science,89.65,,Llama 4 Maverick,Meta,69.8,official,"Instruction-tuned model, non-reasoning (no thinking mode); 0-shot, accuracy",no thinking,https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/MODEL_CARD.md,2025-04-05
|
| "Google, Meta and xAI",MMLU-Pro,Broad expert knowledge,88.07,sample,Llama 4 Maverick,Meta,80.5,official,"Instruction-tuned model, non-reasoning (no thinking mode); 0-shot, macro_avg/acc",no thinking,https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/MODEL_CARD.md,2025-04-05
|
| "Google, Meta and xAI",MMMU,College questions about images,84.67,,Llama 4 Maverick,Meta,73.4,official,"Instruction-tuned model, non-reasoning (no thinking mode); 0-shot, accuracy",no thinking,https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/MODEL_CARD.md,2025-04-05
|
| "Google, Meta and xAI",MathVista,Maths in images,89.1,,Llama 4 Maverick,Meta,73.7,official,"Instruction-tuned model, non-reasoning (no thinking mode); 0-shot, accuracy",no thinking,https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/MODEL_CARD.md,2025-04-05
|
| "Google, Meta and xAI",ChartQA,Reading charts,85.24,thinking off,Llama 4 Maverick,Meta,90.0,official,"Instruction-tuned model, non-reasoning (no thinking mode); 0-shot, relaxed_accuracy",no thinking,https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/MODEL_CARD.md,2025-04-05
|
| "Google, Meta and xAI",DocVQA,Reading documents,94.57,thinking off,Llama 4 Maverick,Meta,94.4,official,"Instruction-tuned model, non-reasoning (no thinking mode); 0-shot, anls, DocVQA (test)",no thinking,https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/MODEL_CARD.md,2025-04-05
|
| "Google, Meta and xAI",MGSM,Maths in 11 languages,85.96,see note,Llama 4 Maverick,Meta,92.3,official,"Instruction-tuned model, non-reasoning (no thinking mode); 0-shot, average/em",no thinking,https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/MODEL_CARD.md,2025-04-05
|
| "Google, Meta and xAI",Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Grok 4.7,xAI,43.1,third-party: Artificial Analysis,"AA harness: HLE text-only subset (2,158 questions), no tools (AA methodology lists no tool use for HLE), 1 repeat; reasoning effort xhigh (AA entry 'Grok 4.7 (xhigh)'), API model grok-4.7",xhigh,https://artificialanalysis.ai/models/grok-4-7,unknown
|
| "Google, Meta and xAI",GPQA Diamond,Graduate-level science,89.65,,Grok 4.3,xAI,90.1,third-party: Artificial Analysis,"AA harness: GPQA Diamond (AA now lists it under 'Legacy Evaluations'); reasoning effort high (AA entry 'Grok 4.3 (high)'), API model grok-4.3",high,https://artificialanalysis.ai/models/grok-4-3,unknown
|
| "Google, Meta and xAI",Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Grok 4.3,xAI,37.2,third-party: Artificial Analysis,"AA harness: HLE text-only subset (2,158 questions), no tools (AA methodology lists no tool use for HLE), 1 repeat; reasoning effort high (AA entry 'Grok 4.3 (high)')",high,https://artificialanalysis.ai/models/grok-4-3,unknown
|
| "Google, Meta and xAI",IFBench,Unseen instructions,73.33,,Grok 4.3,xAI,81.3,third-party: Artificial Analysis,"AA harness: IFBench, no tools, 5 repeats; reasoning effort high (AA entry 'Grok 4.3 (high)')",high,https://artificialanalysis.ai/models/grok-4-3,unknown
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,DeepSeek-V4-Pro,DeepSeek,92.4,official,"thinking on, max reasoning effort, temperature=1.0",max,https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf,2026-09-10
|
| China and India,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,DeepSeek-V4-Pro,DeepSeek,42.7,official,"thinking on, max reasoning effort, temperature=1.0",max,https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813,2026-08-13
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,DeepSeek-V4.1-Flash,DeepSeek,90.9,official,"thinking on, reasoning_effort=100 (Max; = API tier 'max'), temperature=1.0, top_p=0.95 (card) / top_p 1.0 for reasoning evals (tech report 5.3.1)",max,https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash,2026-09-10
|
| China and India,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,DeepSeek-V4.1-Flash,DeepSeek,39.1,official,"thinking on, reasoning_effort=100 (Max; = API tier 'max'), temperature=1.0, top_p=0.95 (card) / top_p 1.0 for reasoning evals (tech report 5.3.1)",max,https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash,2026-09-10
|
| China and India,AIME 2026,"Competition maths, 2026",95.0,,DeepSeek-V4.1-Flash,DeepSeek,100.0,official,"thinking on, reasoning effort 100 (max); temperature 1.0, top_p 1.0 (report sec 5.3.1 reasoning setting); 'Avg Pass@1' per Figure 12; number of samples unstated",max,https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/resolve/main/DeepSeek_V41_Tech_Report.pdf,2026-09-10
|
| China and India,MMLU-Pro,Broad expert knowledge,88.07,sample,DeepSeek-V4-Pro (preview),DeepSeek,87.5,official,"thinking on, Max effort (special system prompt), temperature=1.0, 384K context window",max,https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro,2026-04-24
|
| China and India,LiveCodeBench v6,"Contest coding, 2025 problems",77.1,,DeepSeek-V4-Pro (preview),DeepSeek,93.5,official,"thinking on, Max effort (special system prompt), temperature=1.0, 384K context window; report table labels metric 'Pass@1-COT'",max,https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/89d501aed9/DeepSeek_V4.pdf,2026-05-06
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,Kimi K3,Moonshot AI,93.5,official,thinking always on; reasoning_effort=max; temperature=1.0; top-p=0.95 (single-step tasks); context/max tokens unstated,max,https://huggingface.co/moonshotai/Kimi-K3/raw/main/README.md,2026-09-02
|
| China and India,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Kimi K3,Moonshot AI,46.9,third-party: Artificial Analysis,"AA label ""Kimi K3 (Max)""; reasoning ON, AA effort label ""max""; temperature 0.6 (or lab-recommended); max output tokens = lab-disclosed maximum; zero-shot; 2,158 text-only questions (May 2025 revision, 2,500 total), 1 repeat, open answer graded by equality-checker LLM (GPT-5.6 Luna medium since index v4.1.1), pass@1",max,https://artificialanalysis.ai/models/kimi-k3,unknown (fetched 2026-10-01)
|
| China and India,MMLU-Pro,Broad expert knowledge,88.07,sample,Kimi K3,Moonshot AI,87.97,third-party: Vals AI,"Vals AI run, task ""overall"": 5-shot CoT per category (original-paper prompt); reasoning_effort=max; temperature=unstated; max_output_tokens=262144; provider=Moonshot AI",max,https://www.vals.ai/benchmarks/mmlu_pro,2026-09-01
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,GLM-5.3,Zhipu AI,91.72,third-party: Artificial Analysis,"AA label ""GLM-5.3 (Max)""; reasoning ON, AA effort label ""max""; temperature 0.6 (or lab-recommended); max output tokens = lab-disclosed maximum; zero-shot; 198 questions, 4-option MC, regex extraction, pass@1, zero-shot instruction prompt; GPQA removed from AA Intelligence Index in v4.2 but still run on new releases",max,https://artificialanalysis.ai/models/glm-5-3,unknown (fetched 2026-10-01)
|
| China and India,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,GLM-5.3,Zhipu AI,42.26,third-party: Artificial Analysis,"AA label ""GLM-5.3 (Max)""; reasoning ON, AA effort label ""max""; temperature 0.6 (or lab-recommended); max output tokens = lab-disclosed maximum; zero-shot; 2,158 text-only questions (May 2025 revision, 2,500 total), 1 repeat, open answer graded by equality-checker LLM (GPT-5.6 Luna medium since index v4.1.1), pass@1",max,https://artificialanalysis.ai/models/glm-5-3,unknown (fetched 2026-10-01)
|
| China and India,MMLU-Pro,Broad expert knowledge,88.07,sample,GLM-5.3,Zhipu AI,86.77,third-party: Vals AI,"Vals AI run, task ""overall"": 5-shot CoT per category (original-paper prompt); reasoning_effort=max; temperature=1; max_output_tokens=128000; provider=Zhipu AI",max,https://www.vals.ai/benchmarks/mmlu_pro,2026-09-01
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,MiniMax-M3,MiniMax,92.93,third-party: Artificial Analysis,"AA label ""MiniMax-M3""; reasoning ON, effort unlabelled (AA default config); temperature 0.6 (or lab-recommended); max output tokens = lab-disclosed maximum; zero-shot; 198 questions, 4-option MC, regex extraction, pass@1, zero-shot instruction prompt; GPQA removed from AA Intelligence Index in v4.2 but still run on new releases",,https://artificialanalysis.ai/models/minimax-m3,unknown (fetched 2026-10-01)
|
| China and India,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,MiniMax-M3,MiniMax,38.97,third-party: Artificial Analysis,"AA label ""MiniMax-M3""; reasoning ON, effort unlabelled (AA default config); temperature 0.6 (or lab-recommended); max output tokens = lab-disclosed maximum; zero-shot; 2,158 text-only questions (May 2025 revision, 2,500 total), 1 repeat, open answer graded by equality-checker LLM (GPT-5.6 Luna medium since index v4.1.1), pass@1",,https://artificialanalysis.ai/models/minimax-m3,unknown (fetched 2026-10-01)
|
| China and India,MMLU-Pro,Broad expert knowledge,88.07,sample,MiniMax-M3,MiniMax,84.22,third-party: Vals AI,"Vals AI run, task ""overall"": 5-shot CoT per category (original-paper prompt); reasoning_effort=unstated; temperature=1; max_output_tokens=512000; provider=MiniMax",,https://www.vals.ai/benchmarks/mmlu_pro,2026-09-01
|
| China and India,IFBench,Unseen instructions,73.33,,MiniMax-M3,MiniMax,82.86,third-party: Artificial Analysis,"AA label ""MiniMax-M3""; reasoning ON, effort unlabelled (AA default config); temperature 0.6 (or lab-recommended); max output tokens = lab-disclosed maximum; zero-shot; 294 single-turn questions, 5 repeats, official allenai/IFBench code, LOOSE mode, prompt-level accuracy; standalone eval (removed from index v4.1)",,https://artificialanalysis.ai/models/minimax-m3,unknown (fetched 2026-10-01)
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,Seed 2.0 Pro,ByteDance,88.9,official,"Reasoning/thinking effort unstated (card compares against 'GPT-5.2 High' etc. but gives no Seed effort level, temperature or token budget)",,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,2026-02-14
|
| China and India,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Seed 2.0 Pro,ByteDance,32.4,official,"Reasoning/thinking effort unstated (card compares against 'GPT-5.2 High' etc. but gives no Seed effort level, temperature or token budget); no tools",,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,2026-02-14
|
| China and India,MMLU-Pro,Broad expert knowledge,88.07,sample,Seed 2.0 Pro,ByteDance,87.0,official,"Reasoning/thinking effort unstated (card compares against 'GPT-5.2 High' etc. but gives no Seed effort level, temperature or token budget)",,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,2026-02-14
|
| China and India,AIME 2026,"Competition maths, 2026",95.0,,Seed 2.0 Pro,ByteDance,94.2,official,"Reasoning/thinking effort unstated (card compares against 'GPT-5.2 High' etc. but gives no Seed effort level, temperature or token budget); no tools",,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,2026-02-14
|
| China and India,AIME 2025,"Competition maths, 2025",96.25,,Seed 2.0 Pro,ByteDance,98.3,official,"Reasoning/thinking effort unstated (card compares against 'GPT-5.2 High' etc. but gives no Seed effort level, temperature or token budget); no tools",,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,2026-02-14
|
| China and India,LiveCodeBench v6,"Contest coding, 2025 problems",77.1,,Seed 2.0 Pro,ByteDance,87.8,official,"Reasoning/thinking effort unstated (card compares against 'GPT-5.2 High' etc. but gives no Seed effort level, temperature or token budget)",,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,2026-02-14
|
| China and India,MMMU,College questions about images,84.67,,Seed 2.0 Pro,ByteDance,85.4,official,"Reasoning/thinking effort unstated (card compares against 'GPT-5.2 High' etc. but gives no Seed effort level, temperature or token budget); no tools",,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,2026-02-14
|
| China and India,MathVista,Maths in images,89.1,,Seed 2.0 Pro,ByteDance,89.8,official,"Reasoning/thinking effort unstated (card compares against 'GPT-5.2 High' etc. but gives no Seed effort level, temperature or token budget); no tools",,https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf,2026-02-14
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,Hunyuan Hy4 (preview),Tencent,92.3,official,"thinking on, highest available reasoning setting (card default reasoning_effort 'high'); sampling/temperature for eval unstated",,https://huggingface.co/tencent/Hy4-preview/raw/main/README.md,2026-08-28
|
| China and India,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Hunyuan Hy4 (preview),Tencent,43.4,official,"thinking on, highest available reasoning setting; no tools, text-only subset",,https://huggingface.co/tencent/Hy4-preview/raw/main/README.md,2026-08-28
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,ERNIE 5.1,Baidu,91.0,official,pass@1; thinking/effort unstated,,https://ernie.baidu.com/blog/posts/ernie-5.1-0508-release/,2026-05-09
|
| China and India,AIME 2026,"Competition maths, 2026",95.0,,ERNIE 5.1,Baidu,95.0,official,avg@32; no tools; thinking/effort unstated,,https://ernie.baidu.com/blog/posts/ernie-5.1-0508-release/,2026-05-09
|
| China and India,MMLU-Pro,Broad expert knowledge,88.07,sample,ERNIE 5.1,Baidu,84.3,official,acc; thinking/effort and shots unstated,,https://ernie.baidu.com/blog/posts/ernie-5.1-0508-release/,2026-05-09
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,Step 5 (preview),StepFun,93.5,official,reasoning effort High; sampling unstated,high,https://www.stepfun.com/step-5-preview,unknown (fetched 2026-10-01)
|
| China and India,Humanity's Last Exam (text-only),"Expert questions, text only",38.01,sample,Step 5 (preview),StepFun,46.48,third-party: Artificial Analysis,"AA label ""Step 5 Preview""; reasoning ON, effort unlabelled (AA default config); temperature 0.6 (or lab-recommended); max output tokens = lab-disclosed maximum; zero-shot; 2,158 text-only questions (May 2025 revision, 2,500 total), 1 repeat, open answer graded by equality-checker LLM (GPT-5.6 Luna medium since index v4.1.1), pass@1",,https://artificialanalysis.ai/models/step-5,unknown (fetched 2026-10-01)
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,Sarvam-105B,Sarvam AI,78.7,official,"Reasoning model (thinking on; card examples use enable_thinking=True; mode not stated per benchmark). Card footnote: max context 65,536 tokens; temperature=1.0, top_p=1.0, max_new_tokens=65536.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,AIME 2025,"Competition maths, 2025",96.25,,Sarvam-105B,Sarvam AI,88.3,official,"Reasoning model (thinking on; card examples use enable_thinking=True; mode not stated per benchmark). Card footnote: max context 65,536 tokens; temperature=1.0, top_p=1.0, max_new_tokens=65536. No tools.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,MATH-500,Competition maths,99.4,,Sarvam-105B,Sarvam AI,98.6,official,"Reasoning model (thinking on; card examples use enable_thinking=True; mode not stated per benchmark). Card footnote: max context 65,536 tokens; temperature=1.0, top_p=1.0, max_new_tokens=65536.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,MMLU-Pro,Broad expert knowledge,88.07,sample,Sarvam-105B,Sarvam AI,81.7,official,"Reasoning model (thinking on; card examples use enable_thinking=True; mode not stated per benchmark). Card footnote: max context 65,536 tokens; temperature=1.0, top_p=1.0, max_new_tokens=65536.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,LiveCodeBench v6,"Contest coding, 2025 problems",77.1,,Sarvam-105B,Sarvam AI,71.7,official,"Reasoning model (thinking on; card examples use enable_thinking=True; mode not stated per benchmark). Card footnote: max context 65,536 tokens; temperature=1.0, top_p=1.0, max_new_tokens=65536.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,IFEval,Precise formatting instructions,90.94,,Sarvam-105B,Sarvam AI,84.8,official,"Reasoning model (thinking on; card examples use enable_thinking=True; mode not stated per benchmark). Card footnote: max context 65,536 tokens; temperature=1.0, top_p=1.0, max_new_tokens=65536.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,GPQA Diamond,Graduate-level science,89.65,,Sarvam-30B,Sarvam AI,66.5,official,"Reasoning model (thinking on; mode not stated per benchmark). Card footnote: max context 65,536 tokens; Math500, MMLU, MMLU Pro, GPQA Diamond, AIME 25, HumanEval, MBPP etc. at temperature=1.0, top_p=1.0, max_new_tokens=65536.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,AIME 2025,"Competition maths, 2025",96.25,,Sarvam-30B,Sarvam AI,80.0,official,"Reasoning model (thinking on; mode not stated per benchmark). Card footnote: max context 65,536 tokens; Math500, MMLU, MMLU Pro, GPQA Diamond, AIME 25, HumanEval, MBPP etc. at temperature=1.0, top_p=1.0, max_new_tokens=65536. No tools.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,MATH-500,Competition maths,99.4,,Sarvam-30B,Sarvam AI,97.0,official,"Reasoning model (thinking on; mode not stated per benchmark). Card footnote: max context 65,536 tokens; Math500, MMLU, MMLU Pro, GPQA Diamond, AIME 25, HumanEval, MBPP etc. at temperature=1.0, top_p=1.0, max_new_tokens=65536.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,MMLU-Pro,Broad expert knowledge,88.07,sample,Sarvam-30B,Sarvam AI,80.0,official,"Reasoning model (thinking on; mode not stated per benchmark). Card footnote: max context 65,536 tokens; Math500, MMLU, MMLU Pro, GPQA Diamond, AIME 25, HumanEval, MBPP etc. at temperature=1.0, top_p=1.0, max_new_tokens=65536.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,LiveCodeBench v6,"Contest coding, 2025 problems",77.1,,Sarvam-30B,Sarvam AI,70.0,official,"Reasoning model (thinking on; mode not stated per benchmark). Card footnote: max context 65,536 tokens; Math500, MMLU, MMLU Pro, GPQA Diamond, AIME 25, HumanEval, MBPP etc. at temperature=1.0, top_p=1.0, max_new_tokens=65536.",,https://www.sarvam.ai/blogs/sarvam-30b-105b,2026-03-06
|
| China and India,MILU,11 Indian languages,85.09,sample,Sarvam-30B,Sarvam AI,76.8,official,"Reasoning model (thinking on; mode not stated per benchmark). Card footnote: max context 65,536 tokens; Math500, MMLU, MMLU Pro, GPQA Diamond, AIME 25, HumanEval, MBPP etc. at temperature=1.0, top_p=1.0, max_new_tokens=65536. MILU not listed in any footnote category; shots, language subset and scoring method (log-likelihood vs generative) unstated.",,https://huggingface.co/sarvamai/sarvam-30b,2026-03-23
|
| |
