HealthAgentBench
Agent harnesses completing realistic terminal-based healthcare tasks built from real clinical artifacts.
More about this boardLess
Agentic task success in realistic terminal-based healthcare environments built from real clinical artifacts; evaluates agent harnesses (Claude Code, Codex, Copilot) end to end, not bare models.
Paper: arXiv 2606.31179. The rows are agent harnesses rather than bare models, and the board is run by Microsoft Research; Copilot, Microsoft's own harness, does not top it.
Published by Microsoft Research, released Jul 2026. 54 tasks across 7 environments; 162 trials (3 attempts per task). Mean task success rate, 0-100%, higher better; cost per task also reported.
- Headline metric
- pooled task success rate across 3 attempts x 54 tasks (162 trials)
- Rows on the board
- 12
- Board last updated
- 27 Jul 2026
- Publisher
- Microsoft Research
- Rows
- 12
- Models
- 10 of 12 on the board
- Released
- Jul 2026
- Measured to
- Jul 2026
- Unit
- 54 agentic tasks
- Documents cited
- 1
Every result on HealthAgentBench, on the board's own scale
12 results, one column each, highest first. Mean task success rate, 0-100%, higher better; cost per task also reported.
Column colour is the model's lab. The table below says who ran each result.
Every published result, ranked on this board alone
12 rows. Higher is better. Bars run from 0 to 100; whiskers show a published interval. Select a row for its quote, page and document.
1Claude Code (Opus 5)Anthropic · $3.3/task55Board publisherJul 2026[1]
1 Claude Code (Opus 5) 55% $3.3
2Codex (GPT-5.6-sol)OpenAI · $5.2/task45Board publisherJul 2026[1]
2 Codex (GPT-5.6-sol) 45% $5.2
3Codex (GPT 5.5)OpenAI · $2.8/task42Board publisherJul 2026[1]
3 Codex (GPT 5.5) 42% $2.8
4Copilot (Opus 4.8)Microsoft/Anthropic · $3.1/task36Board publisherJul 2026[1]
4 Copilot (Opus 4.8) 36% $3.1
5Copilot (GPT 5.5)Microsoft/OpenAI · $2.6/task35Board publisherJul 2026[1]
5 Copilot (GPT 5.5) 35% $2.6
6Claude Code (Opus 4.8)Anthropic · $4.0/task32Board publisherJul 2026[1]
6 Claude Code (Opus 4.8) 32% $4.0
7Codex (GPT 5.4)OpenAI · $1.3/task28Board publisherJul 2026[1]
7 Codex (GPT 5.4) 28% $1.3
8Claude Code (Opus 4.7)Anthropic · $4.8/task27Board publisherJul 2026[1]
8 Claude Code (Opus 4.7) 27% $4.8
9Codex (GPT 5.3)OpenAI · $1.0/task22Board publisherJul 2026[1]
Agent | Success Rate | cost/task Codex (GPT 5.3) 22% $1.0
10Claude Code (Opus 4.6)Anthropic · $4.1/task19Board publisherJul 2026[1]
10 Claude Code (Opus 4.6) 19% $4.1
11Claude Code (Sonnet 4.6)Anthropic · $2.9/task17Board publisherJul 2026[1]
11 Claude Code (Sonnet 4.6) 17% $2.9
12Codex (GPT 5.4 Mini)OpenAI · $0.6/task16Board publisherJul 2026[1]
12 Codex (GPT 5.4 Mini) 16% $0.6
1 document supports this board
- 1.HealthAgentBench leaderboardofficial leaderboard, Microsoft Research (HealthAgentBench), 27 Jul 2026. Every result it supports