Gemini 3.8 Flash: Google's cheap model gets serious about coding
Read in : Français

Contents
- What did Google announce on September 2, 2026?
- The numbers that matter
- Is it actually better at coding?
- Does the Cyber version matter to you?
- Why this matters for vibe coders
- What to do now
- If you build in AI Studio or Antigravity
- If your app calls the Gemini API
- Should you switch tools?
- Frequently asked questions
- Frequently asked questions
- Sources
Gemini 3.8 Flash, released on September 2, 2026, is the new version of Google's fast, low-cost model, and Google pitches it as its best model for coding yet. For vibe coders, the news fits in one sentence: a stronger model on long, agentic tasks, at the same introductory price as Gemini 3.7 Flash, but only until the end of the year.
What did Google announce on September 2, 2026?
Google shipped two models built on the same core: Gemini 3.8 Flash for everyone, and Gemini 3.8 Flash Cyber for security work. The official blog post, written by Tulsee Doshi (Google) and Raluca Ada Popa (Google DeepMind), calls 3.8 "our best reasoning and coding model yet, at the same speed and low cost of 3.7."
The pace stands out. Gemini 3.7 Flash came out three weeks earlier, and 9to5Google describes 3.8 as Google's third Flash update in three months.
The numbers that matter
| Item | Gemini 3.8 Flash (announced September 2, 2026) |
|---|---|
| API model ID | gemini-3.8-flash |
| Intro price until Dec 31, 2026 | $0.75 input / $3.75 output per million tokens |
| Price from January 1, 2027 | $1.50 input / $7.50 output per million tokens |
| Context window (Vertex AI) | 1,048,576 tokens |
| Max output (Vertex AI) | 65,536 tokens |
| Thinking levels (Vertex AI) | LOW, MEDIUM (default), HIGH |
| Where to use it | Antigravity, Google AI Studio, Gemini API, Android Studio, Stitch, Gemini app (AI Pro and Ultra) |
Prices and the increase date come from Google's blog post. Context window, output limit and thinking levels come from the model card in the Vertex AI documentation.
Is it actually better at coding?
On paper, yes, and Google leans hard on "agentic" coding, where the model chains steps and tool calls without you. The official post only shows DeepSWE v1.1 (long-horizon software engineering) results as a chart. Android Authority pulled the number: 73.7% for Gemini 3.8 Flash, against 74% for Claude Opus 5, a far more expensive model. The same article says 3.8 Flash beats Gemini 3.7 Flash, Claude Sonnet 5, GPT-5.6 Sol and GPT-5.6 Terra on that test.
These are vendor-reported numbers. A benchmark does not tell you whether the model will understand your half-finished booking app. It tells you that a "Flash" model now plays in the same league as the big ones on this kind of task.
Does the Cyber version matter to you?
Not directly. Gemini 3.8 Flash Cyber is limited to governments, critical infrastructure operators and software maintainers through the Fairwind Program, by application. Google cites two results: the Chrome Security team got 2.6 times more correct patches than with the best commercial models, and on the CWE-Bench patching benchmark the model scored 47.2% against 47.8% for an unnamed frontier model, at a much lower cost.
What does matter to you is one line in the announcement: Google says the 3.8 models make "a significant leap" in prompt injection robustness, as measured by Gray Swan. No figure is given. If your agent reads web pages or emails, that is good news, not a reason to drop your guard.
Why this matters for vibe coders
Three concrete reasons.
- Agents get cheaper to run. A coding agent that works for twenty minutes eats a lot of tokens. A cheap model that holds up on long tasks makes those sessions less expensive, which matters if you pay per use. We break that cost down in what a vibe-coded app really costs.
- Google's stack gets more consistent. The same model lands in Google AI Studio, in the Antigravity agentic editor and in the API. We cover that ecosystem in our Google AI Studio and Antigravity guide.
- The intro price is temporary. From January 1, 2027, the rate doubles. If you are building an app that calls Gemini in production, budget at the 2027 price, not the September one.
What to do now
No need to migrate everything. Here is a simple way to decide.
If you build in AI Studio or Antigravity
- Pick Gemini 3.8 Flash and replay a prompt you know well (a feature you already built, on a copy of the project).
- Compare the output and how many back-and-forths it took.
- Keep the default thinking level, then try a lower one: Google suggests lowering effort to cut token use.
If your app calls the Gemini API
- Switch the model ID on a test branch, not in production. If you have no safety net yet, read Git for vibe coders.
- Measure token usage on 20 typical requests, first with 3.7 Flash, then with 3.8 Flash.
- Recalculate your monthly cost at the January 1, 2027 rate ($1.50 / $7.50).
- Stay on 3.7 Flash if you cannot see a difference: Google says 3.7 Flash "remains fully supported."
Here is a test prompt you can give your agent to compare both models fairly:
You are working on a copy of my project. Add a date filter to the bookings list.
Constraints: only touch the files you need, write a test that checks the filter,
and finish with the exact list of files you changed and how many steps you took.
Run it with each model, then compare three things: does the test pass, how many files changed, and how many tokens were billed.
Should you switch tools?
No. Gemini 3.8 Flash is a model, not a tool. If you are happy in Lovable or Cursor, the only question is whether your tool lets you pick this model. To compare models for your kind of project, see our guide to the best AI models for coding, and for the wider landscape, the complete comparison of vibe coding tools.
Frequently asked questions
Frequently asked questions
Is Gemini 3.8 Flash free?
Not through the API: Google lists $0.75 per million input tokens and $3.75 per million output tokens until December 31, 2026, then $1.50 and $7.50 from January 1, 2027. On the consumer side, the model is in the Gemini app for Google AI Pro and Ultra subscribers. The announcement does not detail any free access in AI Studio, so check there directly.
What is the difference between Gemini 3.8 Flash and 3.7 Flash?
According to Google, 3.8 Flash improves on software engineering, agentic tasks and multi-step reasoning, at the same speed and the same introductory price. The trade-off: it may use more tokens on complex tasks, because it reasons more and calls tools more often. Gemini 3.7 Flash remains fully supported.
Where can I use Gemini 3.8 Flash for vibe coding?
Google lists Antigravity, the Gemini API via Google AI Studio, Android Studio and Stitch for developers, Gemini Enterprise for companies, and the Gemini app, AI Mode in Search and Gemini in Google Sheets for AI Pro and Ultra subscribers. For vibe coding, AI Studio and Antigravity are the two natural entry points.
Can an indie developer get Gemini 3.8 Flash Cyber?
In practice, no. Access goes through the Fairwind Program, aimed at government authorities, critical infrastructure operators and software maintainers, and Google says it vets every applicant. A vibe coder building a SaaS is not the target. The standard model does get improved prompt injection robustness, according to Google.
Can I trust Google's benchmarks?
They are vendor-published results, often shown only as charts, with no independent testing at launch. They show a trend, not a guarantee for your project. The only measurement that counts for you is replaying your own tasks on a copy of your project and comparing output, number of attempts and cost.
Sources
- Introducing Gemini 3.8 Flash and 3.8 Flash Cyber — Google (official blog)
- Gemini 3.8 Flash — Vertex AI documentation, Google Cloud
- Google's Gemini 3.8 Flash is built to "work harder" — Android Authority
- Google Releases Gemini 3.8 Flash and Cyber Variant — Thurrott.com
- Gemini 3.8 Flash rolling out three weeks after last release — 9to5Google


