News and Releases · 2026-09-06 · 1,047 words · 5 min read

Gemini 3.8 Flash released: what changed, from the primary sources

Google's 2 September 2026 release of Gemini 3.8 Flash, dated from the announcement, with the model page limits, the pricing expiry, and the eval table's footnotes.

Gemini model release pricing benchmarks

Google published "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber" on 2 September 2026, signed by Tulsee Doshi and Raluca Ada Popa. It is the primary announcement, and every date and number below comes from it, from the Gemini API model and pricing pages, or from the DeepMind evaluation document the announcement links. Where the evaluation document's own footnotes qualify a number, I say so, because those footnotes are the most useful part of the release for anyone deciding whether to move a workload.

What was released

Two models. Gemini 3.8 Flash, model string gemini-3.8-flash, which the announcement calls its "most intelligent workhorse model, delivering significant improvements from 3.7 Flash across software engineering, agentic tasks, and critical, multi-step reasoning in specialized domains". And Gemini 3.8 Flash Cyber, described as "our most capable cybersecurity model with frontier-level performance in vulnerability detection and automated patching", which is not generally available: access goes through what the post calls the Fairwind Program, to "trusted government authorities, as well as critical infrastructure operators and software maintainers". If you are not one of those, the Cyber model is a press release, not a product.

The general model is available, per the post, in Google AI Studio, the Gemini API, Google Antigravity, Android Studio, Gemini Enterprise, and to consumers through the Gemini app, AI Mode in Search and Google Sheets. The API model page lists it as stable rather than preview.

Limits and capabilities, from the model page

The Gemini API model page for gemini-3.8-flash gives an input limit of 1,048,576 tokens and an output limit of 65,536 tokens. Inputs are text, image, video, audio and PDF; output is text only. It supports function calling, structured outputs, caching, code execution, file search, URL context, search grounding, and thinking at low, medium and high levels; the page states that the "minimal" thinking level returns an error. Computer use is marked preview. Audio and image generation and the Live API are not supported on this model. Batch, Flex and Priority inference tiers are listed.

Two of those lines change how you would use it. The 65,536-token output cap is large enough for long agentic transcripts but is a cap, and the absence of a minimal-thinking mode means the cheapest latency setting on earlier Flash models is not available here.

Pricing, with its expiry date

The pricing page lists Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens, with context caching at $0.075 per million. The same page, and the announcement, state that this is an introductory rate through 31 December 2026, after which the price becomes $1.50 input and $7.50 output ($0.15 caching) from 1 January 2027. The 3.7 Flash and 3.6 Flash rows carry the same terms.

That is a doubling scheduled four months out. A cost forecast built on today's rate for next year's budget is wrong by 2x, and the pricing page says so in a footnote rather than in the headline.

The evaluation table, and what its footnotes say

The announcement links a DeepMind document, "Gemini 3.8 Flash Model evaluation: Approach, methodology and results", whose fourth page is the benchmark table. The columns are Gemini 3.8 Flash, Gemini 3.7 Flash, and four comparison models from two other labs. Selected rows, as printed:

The 3.8-over-3.7 deltas are the cleanest reading, since both columns were run by the same team under the same methodology: +8.4 points on DeepSWE v1.1, +3.6 on Terminal-Bench 2.1, +7.9 on Terminal-Bench 4.0, +8.4 on OSWorld-2.0, +1.3 on HLE-Verified. The pattern the announcement describes, gains concentrated in agentic software engineering and computer use, is what the table shows.

The cross-lab columns need the methodology page. The document states that "all the results for non-Gemini models are sourced from providers' self reported numbers unless otherwise mentioned", and then lists exceptions per benchmark. DeepSWE v1.1 scores come from a public leaderboard, with the Gemini number self-computed on a mini-swe-agent harness at high thinking. Terminal-Bench 2.1 numbers for Gemini are self-computed while the others are taken from the public leaderboard and a third-party aggregator. OSWorld-2.0 runs "were completed before the official OSWorld 2.0 team's 08.08 patch to be compatible with competitor's self-reported numbers". On HLE-Verified, the document notes that "a significant proportion of questions were blocked by content policy filters" for one of the comparison models, which makes that model's 31.0% a measurement of its filters as much as of its reasoning. The document also corrects itself: it originally reported one competitor's DeepSWE score as 74% "due to rounding on the Datacurve public leaderboard".

None of that is unusual for a launch table, and the document deserves credit for writing it down. The practical reading is: trust the within-Gemini deltas, treat the cross-lab columns as a mix of self-reported and self-computed numbers under harnesses that differ per row, and rerun the two or three benchmarks that resemble your workload before believing any column ordering.

What the announcement does not say

No knowledge cutoff appears on the model page as read today. No parameter count, architecture, or training data description is given, which is consistent with earlier Flash releases. The announcement gives a CWE-Bench pass@1 of 47.2% for the Cyber model "compared to a leading frontier model at 47.8%" and a real-world vulnerability discovery success rate "exceeding 70%", without the harness or the sample; those are claims about a model most readers cannot access, and I have not tried to verify them.

What to do

If you run agentic coding or computer-use workloads on 3.7 Flash, the within-family deltas justify an evaluation on your own tasks this month, at a price that is the same as 3.7 Flash today. If you are choosing between labs, the table is not enough on its own; the footnotes tell you why. And whichever you choose, put 1 January 2027 in the budget spreadsheet now.