Gemini 3.8 Flash

by Google / Google DeepMind

The newest Flash tier in the Gemini 3 line, appearing on a DeepSWE leaderboard chart (updated
September 2, 2026) as the “most efficient” model plotted — 74% pass@1 at $2.36 avg cost per task,
high effort.

Part of the Gemini 3 series, following Gemini 3.7 Flash
(shipped August 2026).

What’s confirmed

Sourced from a single DeepSWE benchmark chart (113 tasks, 52/65 configs), not yet corroborated by
yt-base video coverage as of this note:

  • Pass@1: 74% (±1%), high effort setting
  • Avg cost per task: $2.36
  • Output tokens: 143K average
  • Agent steps: 166 average
  • Flagged on the chart as the most cost-efficient point on the whole leaderboard (best score per
    dollar among all plotted models, tying Claude Opus 5’s 74% pass@1 at roughly a
    fifth of Opus 5’s $11.84 avg cost)

Notes

This is a thinly-sourced stub — flag for follow-up once dedicated coverage appears. Given the pace of
the Gemini 3 generation (three Flash releases in 14 weeks between 3.5 and 3.7, per the series note),
3.8 Flash continuing that cadence is unsurprising, but no release-narrative details (pricing changes,
context window, multimodal capability) are confirmed yet beyond the benchmark chart itself.