Google released Gemini 3.8 Flash at $0.75 per million input tokens

Google released Gemini 3.8 Flash on 2 September 2026, alongside a separate Gemini 3.8 Flash Cyber variant aimed at defenders. The post, authored by Tulsee Doshi and Raluca Ada Popa, presents Flash as the general reasoning and coding model and Flash Cyber as a restricted variant for vulnerability discovery and patching.

The price and where to use it

Google says Flash carries an introductory rate of $0.75 per million input tokens and $3.75 per million output tokens, the same as 3.7 Flash. The company adds that the introductory price expires on 31 December 2026, after which it will move to $1.50 and $7.50 per million tokens, so a team using Flash today is paying a rate that will double at the start of the new year.

Availability spans Google AI Studio, the Gemini API, Android Studio, Google Antigravity, Gemini Enterprise, Google Sheets, the Gemini app for Pro and Ultra subscribers, and Google Search AI Mode. The Search AI Mode integration is the one that reaches an audience outside developer tools, because it is the model now generating the conversational answer block for signed-in users on ordinary queries.

What Google says the benchmarks show

Google reports 54.9 percent on HLE-Verified for 3.8 Flash and says the model "outperforms most larger frontier models" on DeepSWE v1.1, the long-horizon software engineering benchmark. On Vals Finance Agent V2 and Harvey's Legal Agent Benchmark the company claims the model beats 3.7 Flash and other leading models but publishes no scores in the post body. Google states the release is "our best reasoning & coding model yet, at the same speed and low cost of 3.7". These are the vendor's own numbers, so worth rerunning against a real workload before rewriting a router.

Flash Cyber is a separate release with a separate audience

The Cyber variant is not open access. Google says it is available to "trusted defenders" through the Fairwind Program and comes with more permissive safety mitigations adjusted for security work. On the internal vulnerability discovery task Google reports success "exceeding 70%" across 20 programming languages, and 47.2 percent pass@1 on the CWE-Bench patching benchmark. The company also says the Cyber variant surpasses 3.5 Flash Cyber on CyberGym without publishing the number.

Two things change for a team already routing production work through a mixed model set. Flash inherits 3.7 Flash's introductory rate, and Google's own claim of matching quality at that price is worth an afternoon of measurement against last month's real traffic before deciding whether to route more work to it. The end-of-year price increase is a real deadline, so a team that uses only Flash in production should budget for the doubled cost from 1 January or write a router that can switch models.

The Search AI Mode change matters more to a marketing team than to a developer. The conversational answer block for signed-in users is now Gemini 3.8 Flash output, and the retrieval and citation behaviour of a new Flash generation is not guaranteed to match the previous one. A brand that watches its position in AI Overviews or AI Mode should rerun its top queries and note whether the citation set has changed before drawing conclusions from the change in tone or length.

Source

Primary source: Introducing Gemini 3.8 Flash and 3.8 Flash Cyber, Google, 2 September 2026.