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

Model

gemini-3.5-flash-lite

7 of 9 tasks passed, in 8,170 ms of wall time, with no thinking tokens reported.

Tasks passed
7 of 9
Total time
8,170 ms
Thinking tokens
none reported
One-word probe
764 ms
TaskResultTime Thinking tokensWhat came back
Four-step arithmetic, number onlyfail682msnot reportedgot "120"
Output the number and nothing elsepass765msnot reportedclean
Reverse a list exactlypass797msnot reportedexact
Answer in exactly seven wordspass906msnot reported7 words
Write a sentence without the letter Efail945msnot reported1 × "e"
Hold three format rules at oncepass945msnot reported3×4, capitalised
Return valid JSON that matches a schemapass1,019msnot reportedpopulation=9776000
Return an array of exactly five itemspass1,047msnot reportedRed,Blue,Green,Yellow,Purple
Find one fact buried in a long documentpass1,064msnot reportedfound

Findings

Where it failed

Write a sentence without the letter E — 1 × "e"
Four-step arithmetic, number only — got "120"

It reports no thinking tokens

Every response came back without a thoughtsTokenCount field. That is not the same as a measured zero: the API simply does not report one here, so cost models that read that field see nothing.

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.