Gemini 3.6 Flash — Model Card and Launch

system card · source date 2026-07-21 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog

Fast, cheap models are where most production agent traffic actually runs, but they get thinner eval treatment than flagships.
Agentic and coding capability is scaffold-dependent, so a number without its harness is close to meaningless.
Safety-framework reporting needs to happen for every release, not only the largest one.

The card documents evaluation across reasoning, coding, agentic, multimodal, and long-context benchmarks: DeepSWE, MLE-Bench, Terminal-Bench 2.1, OSWorld-Verified, GDPval-AA v2, and GDM-MRCR v2.
It includes automated safety, tone, and refusal evaluations plus a Frontier Safety Framework assessment.
Results are reported as deltas versus Gemini 3.5 Flash, with an emphasis on token efficiency rather than raw score.

Token efficiency reported alongside accuracy is the right framing for the Flash class, and it is a reminder that cost-normalized comparison belongs in the card rather than in third-party analysis afterward.
It gives Google/DeepMind a fresh in-window frontier card with detailed agentic tables, and it was immediately re-analyzed across the ecosystem.
Caveat: benchmark numbers are vendor-reported from Google's own harness and scaffolds. Leaderboard-style figures (Terminal-Bench 2.1, GDPval-AA) are harness-dependent and should not be compared cross-lab without matching setups.

Comments

No comments yet.