Gemini 3.8 Flash

Cost and speed evidence

Cost per 1,000 decisions

USD 2.170356

Cost basis

OpenRouter usage.cost receipts

Measured latency · p50 / p95

4.904 s / 8.864 s

Adjusted latency · p50 / p95

4.904 s / 8.864 s

Latency adjustment

none (hosted API, measured as is)

Public accuracy

36.6%

Sealed accuracy

39.7%

Measurement note

Official ImageJev v0.3 A300+P300 pass 1 on 7 Oct 2026 via OpenRouter chat completions (google/gemini-3.8-flash, unchanged v0.2 prompt/parser, max_tokens 384, no reasoning setting). 329 of 600 answers were cut off: the model reasons by default and used the 384-token answer budget before writing its answer (about 10 % in v0.1.5). They count wrong under the unchanged method, so this score mostly reflects that limit, not image understanding. Attempt 1 was stopped before an over-limit image (public items only, not scored). Images are sent with their real MIME type; 7 BMP images were re-encoded losslessly as PNG and 5 images over the 6 MiB request limit were downscaled to 2,048 px on the long side. Provider failures (HTTP errors, timeouts, error bodies) count wrong at cost 0; unparseable answers count wrong at their billed cost. No retries.

Published 2026-10-12. Scores and ranks can change in a later release. See the method notes.