Gemini 3.8 Flash: the same price until New Year's Eve, then double
Google DeepMind shipped Gemini 3.8 Flash on September 2, and the number that matters most is a date. The model keeps the 3.7 Flash introductory price of $0.75 per million input tokens and $3.75 per million output only until December 31, 2026. From January 1, 2027 the price doubles to $1.50 / $7.50. That is the same list price I wrote about for Gemini 3.6 Flash in July, so what Google is really doing is running a four-month discount on the strongest Flash it has ever shipped and telling you in advance when it ends. If your agent pipeline was budgeted on today's price, the cliff has a date, and it is New Year's Eve.
What shipped
The benchmark deltas are large enough to quote even with the usual vendor caveat: DeepSWE v1.1 73.7%, up from 65.3% on 3.7 Flash; Terminal-bench 2.1 89.4%; OSWorld-2.0 59.0%; HLE-Verified 54.9%. Google says it beats Claude Opus 5 on three of those metrics, which for a Flash-tier model at Flash-tier prices is the headline the pricing page is hiding. See /benchmark for how it slots against the rest of the field.
Alongside it, Gemini 3.8 Flash Cyber, a variant tuned for vulnerability detection and auto-patching, is available only through the new Fairwind Program for vetted governments, critical-infrastructure operators and open-source maintainers. The same week OpenAI put its Astra model behind a 'Critical' cyber threshold in its Preparedness Framework. Two labs gating offensive capability behind an access program in the same seven days is a pattern, and one your security team will ask about.
The cost-cliff calculator
Take a concrete agent workload. Say your coding-agent fleet consumes 2 billion input tokens and 200 million output tokens a month, a modest number for a team running agents on every pull request.
| Line item | Until Dec 31, 2026 | From Jan 1, 2027 |
|---|---|---|
| Input, 2B tokens | $1,500 | $3,000 |
| Output, 200M tokens | $750 | $1,500 |
| Monthly bill | $2,250 | $4,500 |
The volumes are mine and illustrative; the prices are Google's. The ratio is what matters: whatever you pay in December, you pay twice in January, with zero change in code or traffic. If you signed a customer contract in October with a margin computed on $0.75, that margin halves in Q1.
A model with an expiry date on its price is a loan, not a discount. Book the repayment before you book the savings.
What to do before the cliff
- Write January's price into the 2027 budget now. Use $1.50 / $7.50 as the planning number for anything running on 3.8 Flash past December.
- Measure output volume, not just input. Output is the expensive axis and the one agent loops inflate. Cut tool chatter and verbose reasoning where the task does not need it; the OpenRouter cost routing pattern of sending only hard steps to the expensive model applies within a single vendor too.
- Route the routine. If a task ran fine on 3.7 Flash, it does not need 3.8 Flash at twice the price. Keep a cheaper tier or a local model for classification, summarization and boilerplate.
- Check /prices monthly. Intro pricing, cache pricing and promo expiries are moving on a calendar now, and Google is not the only vendor doing it.
- Re-run your quality evals on the January routing. Same model, same weights, but if you route differently to save money you need to know what you lost.
The honest limitation
Every benchmark above is Google's own number, and 'beats Opus 5 on three metrics' is Google's choice of which three to publish. Agentic scores have a habit of moving by large margins depending on how much context and reasoning the harness retains between steps, so treat the DeepSWE and Terminal-bench figures as an upper bound until someone outside Google reproduces them. The pricing, on the other hand, is not a claim. It is a schedule, and it is the one part of this release you can plan against with certainty.