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

Task

Reverse a list exactly

No reasoning needed, only careful transcription. Errors here are attention slips, not knowledge gaps.

Measured
2026-09-12 UTC
Result
2 of 2 passed · 4 not asked
Temperature
0
Prompt size
169 chars
ModelResultTimeThinking tokensWhat came back
gemini-3-flash-previewnot asked209msnot reported429 RESOURCE_EXHAUSTED
gemini-2.5-flash-litenot asked232msnot reported429 RESOURCE_EXHAUSTED
gemini-2.5-flashnot asked319msnot reported429 RESOURCE_EXHAUSTED
gemini-3.5-flashnot asked415msnot reported429 RESOURCE_EXHAUSTED
gemini-3.5-flash-litepass797msnot reportedexact
gemini-3.1-flash-litepass1,488msnot reportedexact

What happened

4 models could not be asked

gemini-3-flash-preview returned 429 RESOURCE_EXHAUSTED
gemini-2.5-flash-lite returned 429 RESOURCE_EXHAUSTED
gemini-2.5-flash returned 429 RESOURCE_EXHAUSTED
gemini-3.5-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 2 models we could reach completed this one. The difference is not capability, it is what each spent to get there.

7.1× between fastest and slowest

gemini-3-flash-preview finished in 209ms. gemini-3.1-flash-lite took 1,488ms for the same prompt.

The exact prompt

Reverse the order of this list and output only the reversed list, comma-separated, no spaces after commas: alpha,bravo,charlie,delta,echo,foxtrot,golf,hotel,india,juliet

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.