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

Task

Return an array of exactly five items

Counting its own output while producing structured data — the two failure modes combined.

Measured
2026-09-12 UTC
Result
3 of 3 passed · 3 not asked
Temperature
0
Prompt size
110 chars
ModelResultTimeThinking tokensWhat came back
gemini-3.5-flashnot asked202msnot reported429 RESOURCE_EXHAUSTED
gemini-2.5-flash-litenot asked256msnot reported429 RESOURCE_EXHAUSTED
gemini-2.5-flashnot asked285msnot reported429 RESOURCE_EXHAUSTED
gemini-3.1-flash-litepass762msnot reportedred,blue,green,yellow,purple
gemini-3.5-flash-litepass1,047msnot reportedRed,Blue,Green,Yellow,Purple
gemini-3-flash-previewpass1,491ms143Red,Blue,Green,Yellow,Purple

What happened

3 models could not be asked

gemini-3.5-flash returned 429 RESOURCE_EXHAUSTED
gemini-2.5-flash-lite returned 429 RESOURCE_EXHAUSTED
gemini-2.5-flash 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 3 models we could reach completed this one. The difference is not capability, it is what each spent to get there.

7.4× between fastest and slowest

gemini-3.5-flash finished in 202ms. gemini-3-flash-preview took 1,491ms for the same prompt.

The exact prompt

Return only a JSON array containing exactly 5 strings, each the name of a colour. No markdown, no explanation.

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.