Gemini 3.8 Flash Makes a Fast Model Think Longer

Google has released three Gemini Flash updates in six weeks.[1] The latest, Gemini 3.8 Flash, is built to stay on a problem, call tools repeatedly and spend more of its token budget before stopping.[1]

Forget the usual leaderboard shuffle for a moment. Flash is Google's workhorse tier, aimed at work where speed and price count.[1] If it can take on longer coding jobs, the practical question changes from "Is it smart enough?" to "How much thinking should this particular task buy?"

The price stayed put. The workload did not

Gemini 3.8 Flash launches at the same introductory API rates as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens.[1][2] Those rates are due to double on January 1, 2027.[1][2] The model is already available through the Gemini API and Google AI Studio, as well as products including Android Studio, Gemini Enterprise and the Gemini app for AI Pro and Ultra subscribers.[1][3]

The headline rate does not tell you what a finished task will cost. Google says 3.8 Flash "works harder" on complex jobs by taking extra reasoning steps and making repeated tool calls.[1] It may consume more tokens at higher effort settings, while developers can select lower effort or keep using 3.7 Flash when efficiency matters more.[1]

That caveat is easy to miss in a pricing table. A model can hold its per-token price steady and still make an application more expensive if it reasons for longer, revisits files or calls a search tool several times. The trade can be worthwhile when the extra work prevents a broken patch or an incomplete answer. It is wasteful when the job is extracting a date from a short document.

Developers should watch cost per acceptable result, not only cost per million tokens. That calculation includes retries and tool calls, plus the human review needed after the model finishes.

A Flash model built for jobs, not prompts

Google is positioning 3.8 Flash around long-running coding and agent tasks.[1] In the company's published evaluation, the model reached 54.9% on HLE-Verified and improved over 3.7 Flash on several software-engineering and professional-task benchmarks.[1] Independent coverage of Google's benchmark table notes a mixed picture: 3.8 Flash compares well with larger models on some coding tests, but trails Anthropic's Opus 5 on OSWorld-2.0 computer use and on GDPVal knowledge work.[2]

This is more useful than declaring that a small model beat a big one. Benchmarks isolate particular skills, and vendors choose which results to put in the launch graphic. A coding agent that performs well in a repository may still be clumsy on a desktop or miss that its plan is solving the wrong problem.

For users, the change should show up less as sparkling prose and more as persistence. A longer-running model can inspect more of a codebase, run a test, read the failure, edit the patch and try again. Each loop creates another chance to recover, but also another chance to wander. Good tooling still needs a step limit, a spending cap and a record of what the agent changed.

The same core, with a locked cyber edition

Google also introduced Gemini 3.8 Flash Cyber, which shares the base intelligence but has more permissive cybersecurity controls.[1] It is not generally available.[1] Access runs through a new Fairwind Program for selected government authorities, critical-infrastructure operators and software maintainers.[1]

The company says the cyber model is designed for vulnerability discovery and patching.[1] It reports a 47.2% pass@1 score on Collinear's CWE-Bench, close to an unnamed leading model at 47.8%, plus a success rate above 70% on Google's internal vulnerability benchmark covering 20 programming languages.[1] Those are vendor-reported results, not a promise that the model can safely repair a production system. The internal test is especially hard for outsiders to assess because Google controls both the model and the evaluation.

The access split tells us something about how Google plans to distribute stronger capabilities. The general model is widely available, while the less restricted security variant sits behind an approval process.[1] The New Stack reports that Fairwind includes roughly 650 partners, among them CrowdStrike, Datadog, Palo Alto Networks and Wiz.[2] Two products can grow from the same technical core and still come with different rules about who gets to run them.

What to do before switching

There is no reason to replace every Flash workload on day one. Start with a small set of real tasks and log the whole run. Compare 3.8 with the model already in production on completion rate, elapsed time, tokens consumed, tool calls and reviewer corrections. A lower token price means little if an agent loops twice as long and still needs the same cleanup.

Use effort controls deliberately. Low effort should suit classification, extraction and short transformations. Higher effort has a better case when the task can benefit from testing and revision, such as a multi-file code change. Do not grant broader permissions merely because the model scores better. Keep file, network and deployment access as narrow as the job allows.

Google has made the compromise visible. Gemini 3.8 Flash can spend more computation chasing a better answer, or developers can turn the effort down.[1] That choice belongs in the product design now. "Fast" describes the model family, but it no longer guarantees a short run.

Sources

[1] https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber — Introducing Gemini 3.8 Flash and 3.8 Flash Cyber [2] https://thenewstack.io/google-ships-its-third-gemini-flash-model-in-six-weeks — Google ships its third Gemini Flash model in six weeks [3] https://9to5google.com/2026/09/02/gemini-3-8-flash-launch — Gemini 3.8 Flash rolling out three weeks after last release