TeamBench - Restricted: leaderboard
TeamBench (2026) tests agent coordination under operating-system-enforced role separation on a 90-task stratified subset graded by deterministic scripts, every role filled by the same model at temperature 0. In the restricted condition one agent holds only the Executor permissions and works without the full specification; pass rate in percent, higher is better.
Metric: Pass rate (%) on the 90-task stratified TeamBench leaderboard subset (operating-system-enforced role separation in Docker sandboxes; a run passes when every check of the task's deterministic shell-script grader passes, and a run failing only the attestation check counts as a pass), temperature 0, with one agent holding only the Executor's capabilities (it edits the workspace and runs commands but cannot read the full specification or write the attestation); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 10 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 35.6 | #135 |
| 2 | Claude Opus 4.7 | 33.3 | #85 |
| 3 | Claude Haiku 4.5 | 31.1 | #583 |
| 4 | Claude Sonnet 4.6 | 27.8 | #173 |
| 5 | Gemma 4 31B | 25.6 | #382 |
| 6 | GPT-5.4 Mini | 23.3 | #406 |
| 7 | Gemini 3.1 Pro (Preview) | 22.2 | #99 |
| 8 | Gemini 3.1 Flash Lite | 21.1 | #371 |
| 9 | Gemini 3 Flash | 18.9 | #175 |
| 10 | GPT-OSS-20B | 17.8 | #1183 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=teambench-restricted · How It Works · Data refreshed daily, snapshot 2026-11-21.