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

Task

Four-step arithmetic, number only

Each step is trivial; the chain is where models drop a carry or answer the wrong step.

Measured
2026-09-12 UTC
Result
0 of 2 passed · 4 not asked
Temperature
0
Prompt size
130 chars
ModelResultTimeThinking tokensWhat came back
gemini-3.5-flashnot asked317msnot reported429 RESOURCE_EXHAUSTED
gemini-3-flash-previewnot asked326msnot reported429 RESOURCE_EXHAUSTED
gemini-2.5-flash-litenot asked354msnot reported429 RESOURCE_EXHAUSTED
gemini-2.5-flashnot asked410msnot reported429 RESOURCE_EXHAUSTED
gemini-3.5-flash-litefail682msnot reportedgot "120"
gemini-3.1-flash-litefail1,314msnot reportedgot "188"

What happened

4 models could not be asked

gemini-3.5-flash returned 429 RESOURCE_EXHAUSTED
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. That is a spent quota, not a wrong answer — it is left out of the pass rate rather than counted as a failure.

2 of 2 models failed

gemini-3.5-flash-lite — got "120"
gemini-3.1-flash-lite — got "188"

4.1× between fastest and slowest

gemini-3.5-flash finished in 317ms. gemini-3.1-flash-lite took 1,314ms for the same prompt.

The exact prompt

Compute step by step internally but output only the final number: take 47, multiply by 13, subtract 211, divide by 5, then add 88.

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.