Gemini 3.8 Flash
Made by Google DeepMind. Released 2026-09-02. Accessibility not recorded. Cheapest standard rate we hold: $3.75 per million output tokens, no context split.
The seller has announced a change to $3.75 from 2027-01-01. That is published in advance, not in force.
What it costs
| Tier | Kind | In | Out | Cached in | From |
|---|---|---|---|---|---|
| batch | text | $0.38 | $1.88 | — | seller |
| flex | text | $0.38 | $1.88 | — | seller |
| priority | text | $1.35 | $6.75 | — | seller |
| standard | text | $0.75 | $3.75 | — | seller |
per million tokens, US dollars. * marks a rate from a single registry or two that disagree — the best available number rather than a confirmed one. Every other row is confirmed against a seller or a licensed index.
What it scores, by house
| Benchmark | House | Score | Evaluated | Scaffold |
|---|---|---|---|---|
| chess puzzles | Epoch AI | 0.6±0.05 | 2026-09-02 | high |
| frontiermath tier 4 v2 | Epoch AI | 0.2±0.07 | 2026-09-02 | high |
| frontiermath tiers 1 3 v2 | Epoch AI | 0.7±0.03 | 2026-09-02 | high |
| gpqa diamond | Epoch AI | 1.0±0.01 | 2026-09-02 | high |
| mystery game puzzles | Epoch AI | 0.5±0.05 | 2026-09-02 | high |
| otis mock aime 2024 2025near its ceiling | Epoch AI | 1.0±0.01 | 2026-09-02 | high |
| simpleqa verified | Epoch AI | 0.7±0.01 | 2026-09-02 | high |
| critpt | collected by Epoch AI | 0.1 | — | medium |
| critpt | collected by Epoch AI | 0.2 | — | high |
| critpt | collected by Epoch AI | 0.0 | — | low |
| cursorbench | collected by Epoch AI | 0.7 | — | medium |
| cursorbench | collected by Epoch AI | 0.7 | — | high |
| deepswe | collected by Epoch AI | 0.7 | — | mini-swe-agent |
| proofbench | collected by Epoch AI | 0.5 | — | unknown |
| scicode | collected by Epoch AI | 0.5 | — | medium |
| scicode | collected by Epoch AI | 0.5 | — | low |
| scicode | collected by Epoch AI | 0.5 | — | high |
| webdev arena | collected by Epoch AI | 1567 | — | high |
These scores are not comparable between houses and are not added up. A score is dated by when it was evaluated, not when the model was released, and a score without its scaffold is not a measurement — which is why both are printed.
Headlines that name it
- Gemini 3.8 Flash arrives; Pro-series updates remain paused
- llm-gemini 0.34 adds Gemini 3.8 Flash with three thinking levels
- Google releases Gemini 3.8 Flash and a cybersecurity variant
Matched by the model's name appearing in our headline, literally. There is no link from a story to a model in our data, so this is a search result rather than a claim that the story is about it.
Every rate we hold · Why there is no single best model · Everything from Google