Gemini 3.8 Flash is Google's fourth Flash model in under four months — and every one of them doubles in price on the same day

What shipped

Gemini 3.8 Flash arrived on 2 September 2026. It accepts text, images, audio and video, carries a 1M-token context window with 64K-token text output, and Google reports improvements over 3.7 Flash in software engineering, agentic tasks and multi-step reasoning, with 54.9% on HLE-Verified.

Independent measurement from Artificial Analysis puts it at an Intelligence Index of 59 on high reasoning — up three points from 3.7 Flash — with 57 at medium and 52 at low. Throughput is roughly 300 output tokens per second on high reasoning. Cost per task lands at $0.58 high, $0.41 medium, $0.24 low, and time per task runs 2.5 minutes high to 0.8 minutes low.

It is live for Google AI Pro and Ultra subscribers in the Gemini app, in AI Mode in Google Search and in Gemini for Google Sheets, and for developers through Google Antigravity, AI Studio, the Gemini API and Android Studio.

That is a solid, unremarkable release. The interesting part is the price tag, and specifically what happens when you put it next to the last one.

The same price, and the same deadline

Gemini 3.8 Flash launched at $0.75 per million input tokens and $3.75 per million output, described as introductory and expiring 31 December 2026. From 1 January 2027 the rates become $1.50/$7.50.

Gemini 3.7 Flash launched on 13 August at $0.75/$3.75, described as a 50% introductory cut, expiring 31 December 2026, reverting to $1.50/$7.50.

Those are not similar numbers. They are the same numbers, with the same expiry date, three weeks apart.

The implication is worth stating plainly, because it inverts how introductory pricing normally works. A launch discount usually runs for a fixed window from the launch — adopt later, get the discount later. Here the window is anchored to a calendar date that both models share. Gemini 3.7 Flash buyers got roughly 4.5 months of discounted pricing. Gemini 3.8 Flash buyers get roughly four. The newer model comes with less runway, not more.

So the migration argument that usually accompanies a new release — move to the current version and reset your pricing clock — does not exist here. There is no clock to reset. Whichever Flash version your team standardised on, the bill doubles on 1 January 2027, and adopting 3.8 Flash today does not move that date by a single day.

Four models, one quarter

This is Google’s fourth Flash release in under four months: 3.5 Flash-Lite, then 3.6 Flash with its cyber variant in July, then 3.7 Flash in August, now 3.8 Flash. Each has been a genuine if incremental improvement, and each has cost adopting teams a re-qualification cycle.

Meanwhile the model Google actually promised at I/O in May — the Gemini 3.5 Pro flagship — has now missed its third deadline, with the base model scrapped. The pattern we flagged in August holds: Google keeps shipping the tier it can ship quickly, and the tier it committed to keeps not arriving.

For buyers this produces a specific and underappreciated cost. Version churn at three-week intervals is not free even when every release is an upgrade. Prompt suites drift, evaluation baselines reset, agent scaffolding needs re-testing, and cached-context assumptions change. If you chase every Flash release you will spend more engineering time migrating than the three-point Intelligence Index delta returns.

The defensible policy is to upgrade on your cadence rather than Google’s — quarterly is reasonable — and to break that rule only when a release fixes something you have actually hit in production.

Where the money actually is

Before modelling the January increase, pull your token split. Cached input carries a 90% discount, and for long-context agents that repeatedly read the same codebase or document set, cache reads are where most of the bill lives.

This is the same structural point that ran through Anthropic’s Fable 5.1 cache-read cut this week: frontier pricing is no longer one number you compare across a row. It is a structure of separate rates, and a headline percentage is a claim about the shape of your traffic. A workload that is 80% cache reads will feel the doubling very differently from an output-heavy one, and those two teams should reach different conclusions about whether to move providers.

The broader context is that the floor keeps moving. The open-weight price floor dropped again in late August, DeepSeek’s Flash line continues to undercut on multimodal, and the price war’s direction flipped in August as server costs rose on memory prices. A 2027 doubling is not obviously out of line with where costs are heading — which is precisely why treating it as a bluff would be a mistake.

The Cyber variant

Alongside the standard release, Google shipped Gemini 3.8 Flash Cyber — the same foundation model with more permissive cyber mitigations, tuned toward defensive work. Google reports 86.2% on CyberGym vulnerability discovery and 47.2% pass@1 on CWE-Bench patching, against Claude Fable 5’s 47.8%, and says it prioritised vulnerability fixing over offensive capability.

Access is application-gated through a programme called Fairwind, open to trusted government authorities, critical-infrastructure operators and software maintainers. Pricing for approved users is identical to the standard model.

That makes three restricted tiers announced by three labs inside a week, which we cover separately in the frontier-capability-tiers piece. The short version: the capability you can access is increasingly a question of who you are rather than what you pay.

What to do

Budget the doubling as your base case. Take your current Gemini Flash spend, double the input and output components, leave cached input at its discounted rate, and put that number in your 2027 plan. Treat any extension of the introductory window as upside rather than expectation.

Do not upgrade because a new Flash exists. There is no pricing benefit to moving from 3.7 to 3.8 — the rates and the expiry are identical. Move when a capability gap bites, and batch your migrations.

Instrument your token split now. You cannot evaluate the January increase, or any competing offer, without knowing your fresh-input, cached-input and output proportions. This is a one-afternoon task that will inform every pricing decision you make for the next two quarters.

Keep the exit open. Gemini and Gemini CLI are strong at this price point, and our AI coding tools shortlist covers the alternatives if the 2027 rate changes your calculus. A neutral gateway keeps that switch cheap.

The bottom line

Gemini 3.8 Flash is a good model at a good price, and the price has a date on it that most teams have not written down. The detail that matters is not the three-point capability gain — it is that Google has now launched two consecutive Flash models at identical introductory rates expiring on the same day, which tells you the discount belongs to the calendar rather than to any particular release.

Plan for $1.50/$7.50 from 1 January 2027. If Google extends the window, you will be pleasantly wrong. If you plan for the discount to renew and it does not, you will discover a 100% cost increase in the same week everyone else does.

Update, 3 September 2026 — a cheaper rate appeared the following day, with a different kind of price attached. Meta’s Muse Spark 1.3, released 2 September, lists a contributor endpoint at $0.10/$0.20 per million tokens — well under Gemini 3.8 Flash’s $0.75/$3.75 introductory rate, and under the $1.50/$7.50 this article tells you to budget for from 1 January. It is not a like-for-like substitute for two reasons. The contributor endpoint permits Meta to use your prompts and outputs to improve its products, which Flash’s standard terms do not; and Muse Spark’s own private-data tier is $1.25/$4.25, which is the honest comparison against Flash and lands close to the post-January Gemini rate rather than below it. The planning advice above is unaffected: budget for $1.50/$7.50 in January. But if the January cliff is what pushes you to shop around, note that the cheapest visible number in the market this week is cheap because it is partly paid in data. The arithmetic on Meta’s two tiers.

Questions or corrections? Email Pick Right. Want the full list? See all news.