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

Task

Answer in exactly seven words

A hard format constraint the model must count against, not approximate.

Measured
2026-09-12 UTC
Result
5 of 6 passed
Temperature
0
Prompt size
82 chars
ModelResultTimeThinking tokensWhat came back
gemini-3.5-flash-litepass906msnot reported7 words
gemini-2.5-flash-litepass947msnot reported7 words
gemini-3.1-flash-litepass1,048msnot reported7 words
gemini-2.5-flashfail2,396ms3276 words
gemini-3.5-flashpass2,789ms5557 words
gemini-3-flash-previewpass15,415ms3,9307 words

What happened

1 of 6 models failed

gemini-2.5-flash — 6 words

Thinking more did not mean getting it right

On this task gemini-3-flash-preview spent 3,930 thinking tokens and passed. 3 models passed without reporting a single thinking token, while gemini-2.5-flash spent 327 and got it wrong. Six models is a small sample, so read this as what happened here, not as a law.

17.0× between fastest and slowest

gemini-3.5-flash-lite finished in 906ms. gemini-3-flash-preview took 15,415ms for the same prompt.

The exact prompt

Answer in exactly 7 words, no more and no fewer. Question: what is photosynthesis?

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.