Independent benchmark snapshots

Compare evidence by real work scenario

Scores remain separated by source and metric. They are evidence to inspect, not a universal model ranking.

Filter by scenario

Latest available source results

Snapshots refresh daily. A failed source never overwrites the last valid data. Exact model versions, test setup, and original leaderboard links are retained for each result.

Retrieved: Oct 12, 2026, 1:34 AM · 2 official sources · 663 records

The gold panel marks the top scorer of that official test category, with the gap to the runner-ups shown beside each score. It is not an overall ranking: scores from different benchmarks cannot be combined.

Who leads each test

Rank · Model family · Exact model version · Score

LiveBench

Agentic coding

points
1 DeepSeekdeepseek-v4.1-flash-max 77.3
  1. 2 Claudeclaude-opus-5-5-max-effort −5.6 71.7
  2. 3 qwen3.8-max −12.6 64.6

LiveBench

Coding tasks

points
1 Claudeclaude-sonnet-5-5-max-effort 91.4
  1. 2 ChatGPTgpt-5.6-sol-max −7.4 83.9
  2. 3 smaug-agentic −8.9 82.5

LiveBench

Mathematics

points
1 Claudeclaude-opus-5-5-max-effort 97.1
  1. 2 ChatGPTgpt-6.1-sol-max −0.2 96.8
  2. 3 muse-spark-1.3-xhigh −1.1 95.9

LiveBench

Reasoning

points
1 ChatGPTgpt-6-astra-max 92.7
  1. 2 Claudeclaude-opus-5-5-max-effort −0.5 92.2
  2. 3 Kimikimi-k3 −2.0 90.7

LiveBench

Data analysis

points
1 ChatGPTgpt-6-astra-max 83.0
  1. 2 Claudeclaude-fable-5-max-effort −2.4 80.5
  2. 3 smaug-agentic −3.1 79.9

LiveBench

Instruction following

points
1 Geminigemini-3.8-flash-high 81.4
  1. 2 muse-spark-1.3-xhigh −3.4 78.0
  2. 3 qwen3.8-flash-next −4.3 77.1

LiveBench

Language tasks

points
1 Claudeclaude-fable-5-max-effort 90.7
  1. 2 ChatGPTgpt-6.1-sol-max −0.6 90.1
  2. 3 Geminigemini-3.8-flash-high −2.9 87.8

SWE-bench Verified

Real software-issue fixing

%
1 ClaudeSonar Foundation Agent + Claude 4.5 Opus 79.2%
  1. 2 DoubaoTRAE + Doubao-Seed-Code −0.4 78.8%
  2. 3 Geminilive-SWE-agent + Gemini 3 Pro Preview (2025-11-18) −1.8 77.4%

Every score, side by side

The same comparison with the numbers left in: each cell is that family’s score and rank in that test, and the highest score in each column is highlighted. Counts beside a family name show how many tests it appears in and how many it tops. The dots are tests the family has no public result for — not zeros.

Model family AgenticpointsCodingpointsMathpointsReasoningpointsDatapointsInstructpointsLanguagepointsSWE-bench%
Claude8/84× 71.7#291.4#197.1#192.2#380.5#675.8#690.7#179.2%#1
ChatGPT8/82× 57.3#2183.9#796.8#392.7#183.0#175.6#890.1#274.6%#17
DeepSeek8/81× 77.3#180.0#2295.1#1686.7#3179.5#1371.0#2682.1#2670.0%#46
Gemini8/81× 58.3#1978.9#2993.5#2189.3#1878.5#2381.4#187.8#777.4%#4
GLM8/8 60.9#1579.7#2489.8#3785.8#3576.4#3469.3#3379.9#3172.8%#25
Kimi8/8 62.2#1181.4#1584.4#5690.7#978.7#2171.4#2385.5#1171.2%#35
Doubao1/8 ·······78.8%#3
Qwen1/8 ·······40.6%#137

The podium map across all tests

Each row is a model family, sorted by most first places. A cell shows where that family lands in the test column — only top-3 finishes are marked, and every column keeps its own benchmark’s score.

Model family AgenticCodingMathReasoningDataInstructLanguageSWE-bench
Claude4 3 21122·11
ChatGPT2 3 ·2211·2·
Gemini1 2 ·····133
DeepSeek1 1·······
muse-spark-1.3-xhigh1 1 ··3··2··
Doubao1 ·······2
smaug-agentic2 ·3··3···
Kimi1 ···3····
qwen3.8-flash-next1 ·····3··
qwen3.8-max1 3·······

Full standings by source

Each bar is the best-scoring configuration per model family in that test, so one vendor with many agent setups does not crowd out the others. Expand the original leaderboard for every row.

LiveBench

28 model families compared · Oct 12, 2026, 1:34 AM

Open original leaderboard

LiveBench publishes independently scored, frequently refreshed objective tasks. Category scores shown here are the arithmetic mean of the official per-subtask scores for the selected release, computed locally; each category is displayed separately.

Agentic coding

  1. 1 DeepSeekdeepseek-v4.1-flash-max 77.3
  2. 2 Claudeclaude-opus-5-5-max-effort 71.7−5.6
  3. 3 qwen3.8-max 64.6−12.6
  4. 4 smaug-agentic 64.6−12.6
  5. 5 muse-spark-1.3-xhigh 64.1−13.2
  6. 6 Kimikimi-k3 62.2−15.1
  7. 7 qwen3.8-flash-next 61.6−15.7
  8. 8 qwen3.8-27b 61.4−15.9

Coding tasks

  1. 1 Claudeclaude-sonnet-5-5-max-effort 91.4
  2. 2 ChatGPTgpt-5.6-sol-max 83.9−7.4
  3. 3 smaug-agentic 82.5−8.9
  4. 4 Kimikimi-k3 81.4−9.9
  5. 5 muse-spark-1.3-xhigh 81.1−10.3
  6. 6 DeepSeekdeepseek-v4.1-flash-max 80.0−11.3
  7. 7 GLMglm-5.2 79.7−11.7
  8. 8 Geminigemini-3.7-flash-high 78.9−12.5

Mathematics

  1. 1 Claudeclaude-opus-5-5-max-effort 97.1
  2. 2 ChatGPTgpt-6.1-sol-max 96.8−0.2
  3. 3 muse-spark-1.3-xhigh 95.9−1.1
  4. 4 grok-4.7-xhigh 95.7−1.4
  5. 5 DeepSeekdeepseek-v4-pro-0813 95.1−2.0
  6. 6 mistral-large-4-high 93.6−3.5
  7. 7 Geminigemini-3.7-flash-high 93.5−3.6
  8. 8 grok-4.6 92.6−4.5

Reasoning

  1. 1 ChatGPTgpt-6-astra-max 92.7
  2. 2 Claudeclaude-opus-5-5-max-effort 92.2−0.5
  3. 3 Kimikimi-k3 90.7−2.0
  4. 4 grok-4.6 90.5−2.1
  5. 5 smaug-agentic 90.3−2.4
  6. 6 muse-spark-1.2-xhigh 90.0−2.6
  7. 7 muse-spark-1.3-xhigh 89.7−3.0
  8. 8 Geminigemini-3.8-flash-high 89.3−3.4

Data analysis

  1. 1 ChatGPTgpt-6-astra-max 83.0
  2. 2 Claudeclaude-fable-5-max-effort 80.5−2.4
  3. 3 smaug-agentic 79.9−3.1
  4. 4 muse-spark-1.3-xhigh 79.6−3.4
  5. 5 DeepSeekdeepseek-v4-flash-vision-exp 79.5−3.5
  6. 6 smaug-flash 79.1−3.9
  7. 7 smaug-mini 78.8−4.2
  8. 8 Kimikimi-k3 78.7−4.2

Instruction following

  1. 1 Geminigemini-3.8-flash-high 81.4
  2. 2 muse-spark-1.3-xhigh 78.0−3.4
  3. 3 qwen3.8-flash-next 77.1−4.3
  4. 4 Claudeclaude-fable-5-max-effort 75.8−5.6
  5. 5 ChatGPTgpt-6-astra-max 75.6−5.8
  6. 6 grok-4.7-xhigh 75.3−6.1
  7. 7 muse-spark-1.2-xhigh 74.3−7.1
  8. 8 qwen3.8-max 74.1−7.3

Language tasks

  1. 1 Claudeclaude-fable-5-max-effort 90.7
  2. 2 ChatGPTgpt-6.1-sol-max 90.1−0.6
  3. 3 Geminigemini-3.8-flash-high 87.8−2.9
  4. 4 Kimikimi-k3 85.5−5.2
  5. 5 smaug-agentic 84.4−6.3
  6. 6 grok-4.6 83.7−7.0
  7. 7 grok-4.5 82.8−7.9
  8. 8 muse-spark-1.3-xhigh 82.8−7.9

SWE-bench Verified

87 model families compared · Oct 12, 2026, 1:34 AM

Open original leaderboard

This result concerns an agent configuration and repository issue resolution, not generic chat coding.

Real software-issue fixing

  1. 1 ClaudeSonar Foundation Agent + Claude 4.5 Opus 79.2%
  2. 2 DoubaoTRAE + Doubao-Seed-Code 78.8%−0.4
  3. 3 Geminilive-SWE-agent + Gemini 3 Pro Preview (2025-11-18) 77.4%−1.8
  4. 4 Atlassian Rovo Dev (2025-09-02) 76.8%−2.4
  5. 5 ACoder 76.4%−2.8
  6. 6 MiniMax M2.5 (high) 75.8%−3.4
  7. 7 Warp 75.6%−3.6
  8. 8 Harness AI 74.8%−4.4

Refresh notes: arena/text: 429 https://api.wulong.dev/arena-ai-leaderboards/v1/leaderboard?name=text | arena/search: 429 https://api.wulong.dev/arena-ai-leaderboards/v1/leaderboard?name=search | arena/code: 429 https://api.wulong.dev/arena-ai-leaderboards/v1/leaderboard?name=code

How to read this page

LiveBench reports category scores on objective tasks. Arena reports blind human-preference Elo. SWE-bench Verified reports the rate at which an agent configuration resolves real repository issues.

Important limits

Scores from different benchmarks cannot be added together. The everyday-life view uses a preference proxy because no general authoritative benchmark covers that scenario.

Add to Chrome — free