Google Gemini 3.8 Flash: third Flash version in six weeks
Google released Gemini 3.8 Flash on September 2, 2026 — the third Flash version within six weeks. Per Google’s model card, the model is based on Gemini 3.7 Flash and is not a new base model but a tuning step: optimized for long-horizon coding and autonomous agents. The price stays identical to its predecessor — with an announced doubling on January 1, 2027. With this, Google keeps the pace high in the Flash line and continues its positioning of offering near-frontier performance at a volume price.
What Google states at launch
- Released September 2, 2026 — third Flash version in six weeks, based on Gemini 3.7 Flash.
- Price: 0.75 US dollars per million input tokens, 3.75 US dollars per million output tokens — identical to 3.7 Flash.
- Price doubling announced: both values rise on January 1, 2027.
- Context: 1M token input, 64K output; inputs as text, image, audio, video, and PDF.
- Focus: tuned for long-horizon coding and autonomous agents.
- Benchmark claim (Google): beats 3.7 Flash on all published benchmarks and Claude Opus 5 on three.
Where things stood
The Flash line is Google’s fast, cheap model series — the workhorse for coding, agents, and high volume. Over recent months, Google noticeably raised the cadence: Gemini 3.7 Flash was followed by 3.8 Flash within a few weeks. Already with the announcement that Flash becomes the default (see our news on Gemini 3.5 Pro), Google made clear that the fast line, not the Pro model, is meant to carry everyday work.
Price was the central selling point: 0.75 US dollars input and 3.75 US dollars output per million tokens sit well below competitors’ frontier models. For agent workflows that consume many tokens, Flash was thus the low-cost default choice for months.
What is new
1. A tuning step, not a new base model. Google itself describes Gemini 3.8 Flash as based on 3.7 Flash. That is a relevant framing: this is about targeted improvements in coding and agent behavior, not an architectural leap. If you already run 3.7 Flash, treat the switch as an incremental update.
2. The price holds — for now. At launch, 3.8 Flash costs exactly the same as 3.7 Flash. Important for planning: Google has already announced the doubling of both values on January 1, 2027. If you plan budgets across the new year, calculate with the higher price.
3. Benchmark lead as a Google claim. Per Google, 3.8 Flash beats its predecessor on all published benchmarks and Claude Opus 5 on three. These are vendor figures, not independent tests — and comparing a fast, cheap model against a top-tier model on selected benchmarks should be read with caution. For your own selection, what counts is testing on real tasks.
4. A Cyber variant for trusted testers. In parallel, Google brings Gemini 3.8 Flash Cyber (replacing 3.5) via the Fairwind program to selected testers — focused on autonomous vulnerability discovery. Regular 3.8 Flash is available in the Gemini app (Pro/Ultra), in AI Mode, in Sheets, Antigravity, AI Studio, and via the Gemini API.
Reading
The real story is the cadence. Three Flash versions in six weeks is an unusually tight release window — and it says more about the competitive situation than about a technical breakthrough. Google keeps attention on the Flash line at high frequency and signals it is playing at the front of the race for agent and coding budgets. The impression: this is about continuous, visible motion, not a single big leap.
For practice, the announced price doubling is the more important data point than the benchmark claim. A cheap model whose price jumps to double at the new year noticeably changes the math for token-heavy workflows. Betting on 3.8 Flash today wins a short-term price advantage — but you should base your calculation on the price from January 2027, not the current one.
What you can do now
If you already use Flash: treat the switch to 3.8 as an incremental update and test on your own coding and agent tasks whether the improvements are measurable for you.
If you plan budgets: calculate with the price doubling on January 1, 2027. A cost model that only works at the current price does not hold across the new year.
If you compare models: take the benchmark statements as a Google claim and test against your current setup. A lead on three selected benchmarks is not general superiority.