Rankings / Anthropic
Claude Sonnet 4.5 20250929
released 2025-09-29
BenchAtlas Index
as of 2026-07-14
thinking
rank #69
10 families · 4 categories · medium
32K budget
rank #111
6 families · 3 categories · medium
59K budget
rank #118
5 families · 3 categories · medium
Benchmark evidence
112 results
external indices
| ECI | 8K budget | 146.8 points | independent | Epoch AI Benchmarking Hub |
| ECI | 12K budget | 146.8 points | independent | Epoch AI Benchmarking Hub |
| ECI | 2K budget | 146.8 points | independent | Epoch AI Benchmarking Hub |
| ECI | 1K budget | 146.8 points | independent | Epoch AI Benchmarking Hub |
| ECI | 16K budget | 146.8 points | independent | Epoch AI Benchmarking Hub |
| ECI | 146.8 points | independent | Epoch AI Benchmarking Hub | |
| ECI | 32K budget | 146.8 points | independent | Epoch AI Benchmarking Hub |
| ECI | 59K budget | 146.8 points | independent | Epoch AI Benchmarking Hub |
factuality
| MASK contamination=Potential contamination warning: This model was evaluated after the public release of MASK, allowing model builder access to the prompts and solutions. | thinking | 96.1 % 95.6–96.7 | independent | Scale Labs |
| MASK contamination=Potential contamination warning: This model was evaluated after the public release of MASK, allowing model builder access to the prompts and solutions. | 86.4 % 84.9–87.9 | independent | Scale Labs | |
| SimpleQA Verified | 59K budget | 23.6 % | independent | Epoch AI Benchmarking Hub 2025-12-09 |
| SimpleQA Verified | 13.0 % | independent | Epoch AI Benchmarking Hub 2025-12-08 |
human preference
| Coding (style control)older version arena=text · category=coding · style_control=true | 1513.2 1508.0–1518.4 | community | LMArena Leaderboard Dataset | |
| Industry Software And It Services (style control)older version arena=text · category=industry_software_and_it_services · style_control=true | 1497.9 1493.6–1502.2 | community | LMArena Leaderboard Dataset | |
| Hard Prompts English (style control)older version arena=text · category=hard_prompts_english · style_control=true | 1490.6 1485.8–1495.4 | community | LMArena Leaderboard Dataset | |
| Expert (style control)older version arena=text · category=expert · style_control=true | 1485.6 1477.2–1494.1 | community | LMArena Leaderboard Dataset | |
| Longer Query (style control)older version arena=text · category=longer_query · style_control=true | 1483.7 1478.8–1488.5 | community | LMArena Leaderboard Dataset | |
| Hard Prompts (style control)older version arena=text · category=hard_prompts · style_control=true | 1483.5 1479.8–1487.3 | community | LMArena Leaderboard Dataset | |
| Chinese (style control)older version arena=text · category=chinese · style_control=true | 1480.7 1470.4–1491.0 | community | LMArena Leaderboard Dataset | |
| Multi Turn (style control)older version arena=text · category=multi_turn · style_control=true | 1479.1 1473.3–1484.8 | community | LMArena Leaderboard Dataset | |
| French (style control)older version arena=text · category=french · style_control=true | 1477.0 1461.0–1493.0 | community | LMArena Leaderboard Dataset | |
| Industry Life And Physical And Social Science (style control)older version arena=text · category=industry_life_and_physical_and_social_science · style_control=true | 1472.9 1467.1–1478.8 | community | LMArena Leaderboard Dataset | |
| Industry Medicine And Healthcare (style control)older version arena=text · category=industry_medicine_and_healthcare · style_control=true | 1472.8 1463.7–1481.8 | community | LMArena Leaderboard Dataset | |
| English (style control)older version arena=text · category=english · style_control=true | 1468.6 1464.7–1472.5 | community | LMArena Leaderboard Dataset | |
| Spanish (style control)older version arena=text · category=spanish · style_control=true | 1465.3 1451.2–1479.3 | community | LMArena Leaderboard Dataset | |
| Industry Legal And Government (style control)older version arena=text · category=industry_legal_and_government · style_control=true | 1463.5 1455.2–1471.9 | community | LMArena Leaderboard Dataset | |
| Instruction Following (style control)older version arena=text · category=instruction_following · style_control=true | 1463.4 1458.6–1468.1 | community | LMArena Leaderboard Dataset | |
| Industry Business And Management And Financial Operations (style control)older version arena=text · category=industry_business_and_management_and_financial_operations · style_control=true | 1460.7 1455.2–1466.3 | community | LMArena Leaderboard Dataset | |
| Exclude Ties (style control)older version arena=text · category=exclude_ties · style_control=true | 1458.8 1454.9–1462.8 | community | LMArena Leaderboard Dataset | |
| Overall (style control) arena=text · category=overall · style_control=true | 1455.2 1452.3–1458.1 | community | LMArena Leaderboard Dataset | |
| Industry Writing And Literature And Language (style control)older version arena=text · category=industry_writing_and_literature_and_language · style_control=true | 1454.2 1449.1–1459.4 | community | LMArena Leaderboard Dataset | |
| Creative Writing (style control)older version arena=text · category=creative_writing · style_control=true | 1453.9 1447.7–1460.1 | community | LMArena Leaderboard Dataset | |
| Russian (style control)older version arena=text · category=russian · style_control=true | 1453.8 1446.6–1461.0 | community | LMArena Leaderboard Dataset | |
| Industry Mathematical (style control)older version arena=text · category=industry_mathematical · style_control=true | 1452.7 1442.8–1462.5 | community | LMArena Leaderboard Dataset | |
| German (style control)older version arena=text · category=german · style_control=true | 1444.7 1427.4–1462.0 | community | LMArena Leaderboard Dataset | |
| Industry Entertainment And Sports And Media (style control)older version arena=text · category=industry_entertainment_and_sports_and_media · style_control=true | 1442.6 1437.1–1448.2 | community | LMArena Leaderboard Dataset | |
| Non English (style control)older version arena=text · category=non_english · style_control=true | 1439.2 1435.5–1442.9 | community | LMArena Leaderboard Dataset | |
| Polish (style control)older version arena=text · category=polish · style_control=true | 1438.2 1425.3–1451.1 | community | LMArena Leaderboard Dataset | |
| Math (style control)older version arena=text · category=math · style_control=true | 1427.9 1419.3–1436.5 | community | LMArena Leaderboard Dataset | |
| Korean (style control)older version arena=text · category=korean · style_control=true | 1414.9 1396.8–1432.9 | community | LMArena Leaderboard Dataset | |
| Japanese (style control)older version arena=text · category=japanese · style_control=true | 1408.6 1385.0–1432.2 | community | LMArena Leaderboard Dataset |
knowledge science
| GPQA Diamond | 59K budget | 82.3 % | independent | Epoch AI Benchmarking Hub 2025-10-28 |
| GPQA Diamond | 32K budget | 81.7 % | independent | Epoch AI Benchmarking Hub 2025-10-21 |
| GPQA Diamond implementation=vals-ai | thinking | 81.6 % | independent | Vals AI |
| GPQA Diamond | 16K budget | 78.8 % | independent | Epoch AI Benchmarking Hub 2025-10-28 |
| GPQA Diamond | 73.7 % | independent | Epoch AI Benchmarking Hub 2025-09-29 | |
| Humanity's Last Exam contamination=Potential contamination warning: This model was evaluated after the public release of HLE, allowing model builder access to the prompts and solutions. · implementation=scale | thinking | 13.7 % 12.4–15.1 | independent | Scale Labs |
| Humanity's Last Exam contamination=Potential contamination warning: This model was evaluated after the public release of HLE, allowing model builder access to the prompts and solutions. · implementation=scale | 7.5 % 6.5–8.6 | independent | Scale Labs | |
| MMLU-Pro implementation=vals-ai | thinking | 87.4 % | independent | Vals AI |
long context instruction
| MultiChallenge | thinking | 55.3 % 53.9–56.8 | independent | Scale Labs |
multimodal
| MMMU implementation=vals-ai | thinking | 79.3 % | independent | Vals AI |
| VISTA | thinking | 48.8 % 47.1–50.4 | independent | Scale Labs |
| VISTA | 45.0 % 44.9–45.1 | independent | Scale Labs |
professional
| CaseLaw | thinking | 62.2 % | independent | Vals AI |
| CorpFin | thinking | 62.0 % | independent | Vals AI |
| CorpFin | 60.8 % | independent | Vals AI | |
| LegalBench | thinking | 84.1 % | independent | Vals AI |
| MedQA | thinking | 94.7 % | independent | Vals AI |
| TaxEval | thinking | 73.3 % | independent | Vals AI |
reasoning math
| AIME implementation=vals-ai | thinking | 88.2 % | independent | Vals AI |
| ARC-AGI-1older version split=public_eval · model_type=CoT | thinking · 32K budget | 73.8 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | thinking · 32K budget | 63.7 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | thinking · 16K budget | 63.6 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | thinking · 8K budget | 53.5 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | thinking · 16K budget | 48.3 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | thinking · 8K budget | 46.5 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | thinking · 1K budget | 36.6 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=Base LLM | 35.4 % | independent | ARC Prize Leaderboard | |
| ARC-AGI-1older version split=semi_private · model_type=CoT | thinking · 1K budget | 31.0 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=Base LLM | 25.5 % | independent | ARC Prize Leaderboard | |
| ARC-AGI-2 split=public_eval · model_type=CoT | thinking · 32K budget | 14.7 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | thinking · 32K budget | 13.6 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | thinking · 16K budget | 6.9 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | thinking · 8K budget | 6.9 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | thinking · 16K budget | 6.5 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | thinking · 1K budget | 5.8 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | thinking · 8K budget | 4.9 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=Base LLM | 3.8 % | independent | ARC Prize Leaderboard | |
| ARC-AGI-2 split=public_eval · model_type=Base LLM | 3.8 % | independent | ARC Prize Leaderboard | |
| ARC-AGI-2 split=public_eval · model_type=CoT | thinking · 1K budget | 2.1 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | thinking · 32K budget | 0.8 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | thinking · 32K budget | 0.8 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | thinking · 32K budget | 0.5 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | thinking · 32K budget | 0.4 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | thinking · 16K budget | 0.4 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | thinking · 16K budget | 0.3 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | thinking · 16K budget | 0.3 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | thinking · 16K budget | 0.2 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | thinking · 8K budget | 0.2 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | thinking · 8K budget | 0.2 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | thinking · 8K budget | 0.2 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | thinking · 8K budget | 0.2 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | thinking · 1K budget | 0.1 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | thinking · 1K budget | 0.1 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=Base LLM | 0.1 usd_per_task | independent | ARC Prize Leaderboard | |
| ARC-AGI-2 split=semi_private · model_type=Base LLM | 0.1 usd_per_task | independent | ARC Prize Leaderboard | |
| ARC-AGI-1older version split=semi_private · model_type=CoT | thinking · 1K budget | 0.1 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | thinking · 1K budget | 0.1 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=Base LLM | 0.1 usd_per_task | independent | ARC Prize Leaderboard | |
| ARC-AGI-1older version split=public_eval · model_type=Base LLM | 0.1 usd_per_task | independent | ARC Prize Leaderboard | |
| EnigmaEval contamination=Potential contamination warning: This model was evaluated after the public release of EnigmaEval, allowing model builder access to the prompts and solutions. | thinking | 6.0 % 4.7–7.3 | independent | Scale Labs |
| EnigmaEval contamination=Potential contamination warning: This model was evaluated after the public release of EnigmaEval, allowing model builder access to the prompts and solutions. | 3.4 % 2.4–4.4 | independent | Scale Labs | |
| FrontierMath | 32K budget | 15.2 % | independent | Epoch AI Benchmarking Hub 2025-11-16 |
| FrontierMath | 59K budget | 13.5 % | independent | Epoch AI Benchmarking Hub 2025-11-13 |
| FrontierMath | 9.3 % | independent | Epoch AI Benchmarking Hub 2025-11-16 | |
| FrontierMath Tier 4older version | 32K budget | 4.2 % | independent | Epoch AI Benchmarking Hub 2025-10-22 |
| FrontierMath Tier 4older version | 2.1 % | independent | Epoch AI Benchmarking Hub 2025-09-29 | |
| MATH Level 5 | 32K budget | 97.7 % | independent | Epoch AI Benchmarking Hub 2025-10-21 |
| OTIS Mock AIME 2024–2025 | 59K budget | 77.8 % | independent | Epoch AI Benchmarking Hub 2025-10-28 |
| OTIS Mock AIME 2024–2025 | 32K budget | 77.8 % | independent | Epoch AI Benchmarking Hub 2025-10-21 |
| OTIS Mock AIME 2024–2025 | 16K budget | 71.1 % | independent | Epoch AI Benchmarking Hub 2025-10-28 |
| OTIS Mock AIME 2024–2025 | 35.6 % | independent | Epoch AI Benchmarking Hub 2025-09-29 |
Agent + model results
systems, not bare-model scores
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench bash-only | 71.4 % | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench Verified | 71.4 % | unverified | SWE-bench Leaderboard |
| agent + model Epoch Inspect harness + Claude Sonnet 4.5 20250929 | SWE-bench Verified | 71.3 % | independent | Epoch AI Benchmarking Hub |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench Verified | 70.6 % | community | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench bash-only | 70.6 % | community | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench Multilingual | 67.0 % | community | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench Multilingual | 0.7 usd_per_task | community | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench bash-only | 0.7 usd_per_task | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench Verified | 0.7 usd_per_task | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench Verified | 0.6 usd_per_task | community | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | SWE-bench bash-only | 0.6 usd_per_task | community | SWE-bench Leaderboard |
| agent + model terminus-2 + Claude Sonnet 4.5 20250929 | Terminal-Bench 2.0 | 42.8 % | community | Terminal-Bench Leaderboard |
| agent + model OpenHands + Claude Sonnet 4.5 20250929 | Terminal-Bench 2.0 | 42.6 % | community | Terminal-Bench Leaderboard |
| agent + model mini-SWE-agent + Claude Sonnet 4.5 20250929 | Terminal-Bench 2.0 | 42.5 % | community | Terminal-Bench Leaderboard |
| agent + model Claude Code + Claude Sonnet 4.5 20250929 | Terminal-Bench 2.0 | 40.1 % | community | Terminal-Bench Leaderboard |
These scores measure the whole agent system (scaffold, tools, budgets) — they are never merged into the bare model’s numbers.