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

Task

Output the number and nothing else

Models like to explain. Here any extra character is a failure a parser would hit.

Measured
2026-09-12 UTC
Result
5 of 5 passed · 1 not asked
Temperature
0
Prompt size
87 chars
ModelResultTimeThinking tokensWhat came back
gemini-3-flash-previewnot asked224msnot reported429 RESOURCE_EXHAUSTED
gemini-2.5-flash-litepass721msnot reportedclean
gemini-3.5-flash-litepass765msnot reportedclean
gemini-2.5-flashpass1,315ms118clean
gemini-3.5-flashpass1,499ms155clean
gemini-3.1-flash-litepass1,729msnot reportedclean

What happened

1 model could not be asked

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 5 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 155 thinking tokens and passed. 3 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.

7.7× between fastest and slowest

gemini-3-flash-preview finished in 224ms. gemini-3.1-flash-lite took 1,729ms for the same prompt.

The exact prompt

Output only the number, with no words, no punctuation, no explanation. What is 17 * 23?

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.