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

Model

gemini-3-flash-preview

5 of 5 tasks passed, in 35,298 ms of wall time and 8,493 thinking tokens.

Tasks passed
5 of 5
Total time
35,298 ms
Thinking tokens
8,493
One-word probe
1,130 ms
Not asked
4 (quota)
TaskResultTime Thinking tokensWhat came back
Return an array of exactly five itemspass1,491ms143Red,Blue,Green,Yellow,Purple
Return valid JSON that matches a schemapass2,147ms228population=9776000
Find one fact buried in a long documentpass2,376ms262found
Write a sentence without the letter Epass13,869ms3,9306 words, clean
Answer in exactly seven wordspass15,415ms3,9307 words

Findings

It passed everything we could ask

All 5 tasks completed. What separates it from the others here is cost, not capability.

It reports thinking tokens

8,493 across 5 tasks. Those are billed, and they land before the first output token — which is why wall time and output length disagree on this model.

4 tasks could not be asked

The free-tier quota was spent before we got to Output the number and nothing else, Hold three format rules at once, Reverse a list exactly, Four-step arithmetic, number only. Those are left out of the pass rate rather than counted against the model.

How this was measured

Every task was sent at temperature 0, one request at a time, and judged by code — the checkers live in bench/tasks.mjs. Failed responses are stored verbatim so the verdict can be re-read. See all models and tasks.