osmosMind

Where the models stand.

We build and maintain living benchmarks across terminal ops, software engineering, reasoning, health, and agents — then publish exactly how each one is constructed. Open a benchmark for the full ranking and the methodology behind it.

Terminal-Bench

Real command-line tasks — diagnosis, repair, and multi-step system operations verified in a clean environment.

View leaderboard

Top 5Resolved %
  1. 1AnthropicClaude Opus 4.858.2
  2. 2OpenAIGPT-5.454.7
  3. 3GoogleGemini 3 Pro51.3
  4. 4AnthropicClaude Sonnet 4.649.8
  5. 5DeepSeekDeepSeek V444.1

SWE-Bench Verified

Real GitHub issues fixed in runnable repositories, with each patch judged pass/fail by the project's own tests.

View leaderboard

Top 5Resolved %
  1. 1AnthropicClaude Opus 4.879.4
  2. 2OpenAIGPT-5.476.1
  3. 3AnthropicClaude Sonnet 4.674.8
  4. 4GoogleGemini 3 Pro71.5
  5. 5DeepSeekDeepSeek V466.2

STEM Reasoning

Long-chain math, science, and multi-step derivation problems graded against reference answers and rubrics.

View leaderboard

Top 5Accuracy %
  1. 1OpenAIGPT-5.488.9
  2. 2AnthropicClaude Opus 4.887.6
  3. 3GoogleGemini 3 Pro86.2
  4. 4DeepSeekDeepSeek V481.0
  5. 5GrokGrok 479.1

health_bench

Clinical and health-knowledge tasks scored by physician-written rubrics with explicit safety boundaries.

View leaderboard

Top 5Score
  1. 1AnthropicClaude Opus 4.871.2
  2. 2OpenAIGPT-5.469.8
  3. 3GoogleGemini 3 Pro67.1
  4. 4AnthropicClaude Sonnet 4.664.5
  5. 5DeepSeekDeepSeek V460.3

Agent-Ops

Long-horizon agent sessions over real tools — measured end-to-end, including failure recovery.

View leaderboard

Top 5Success %
  1. 1AnthropicClaude Opus 4.864.7
  2. 2OpenAIGPT-5.461.2
  3. 3GoogleGemini 3 Pro57.8
  4. 4AnthropicClaude Sonnet 4.655.1
  5. 5DeepSeekDeepSeek V448.6