Agent BenchWe give models the same task and publish what it cost them.

Task

Hold three format rules at once

Each rule is easy alone. Together they need the model to check its own output before finishing.

Measured
2026-09-12 UTC
Result
4 of 4 passed · 2 not asked
Temperature
0
Prompt size
137 chars
ModelResultTimeThinking tokensWhat came back
gemini-2.5-flash-litenot asked200msnot reported429 RESOURCE_EXHAUSTED
gemini-3-flash-previewnot asked213msnot reported429 RESOURCE_EXHAUSTED
gemini-3.1-flash-litepass782msnot reported3×4, capitalised
gemini-3.5-flash-litepass945msnot reported3×4, capitalised
gemini-2.5-flashpass2,262ms3263×4, capitalised
gemini-3.5-flashpass4,199ms9313×4, capitalised

What happened

2 models could not be asked

gemini-2.5-flash-lite returned 429 RESOURCE_EXHAUSTED
gemini-3-flash-preview returned 429 RESOURCE_EXHAUSTED. That is a spent quota, not a wrong answer — it is left out of the pass rate rather than counted as a failure.

Every model passed

All 4 models we could reach completed this one. The difference is not capability, it is what each spent to get there.

Thinking more did not mean getting it right

On this task gemini-3.5-flash spent 931 thinking tokens and passed. 2 models passed without reporting a single thinking token. Six models is a small sample, so read this as what happened here, not as a law.

21.0× between fastest and slowest

gemini-2.5-flash-lite finished in 200ms. gemini-3.5-flash took 4,199ms for the same prompt.

The exact prompt

Write exactly 3 lines about rain. Each line must be exactly 4 words. Each line must start with a capital letter. Output only the 3 lines.

Pass/fail is decided by code, not by reading the answer. The check for this task is in bench/tasks.mjs, and failed responses are stored verbatim so the verdict can be re-read.