Gemini 3.8 Flash Ships With a Price That Doubles January 1

$0.75 per million input tokens, $3.75 per million output. That is what Google is charging for Gemini 3.8 Flash, released September 2. The price comes with an expiry date attached: December 31, 2026. On January 1 the same model costs $1.50 and $7.50, exactly double.

Google shipped a second version the same day, Gemini 3.8 Flash Cyber. Same core model, a completely different door to get in through.

What the published numbers show

Google's own announcement puts 3.8 Flash ahead of the previous Flash generation on coding and tool use. The figures it published: 54.9% on HLE-Verified, 47.2% Pass@1 on CWE-Bench, and a better than 70% success rate at discovering real-world vulnerabilities across 20 programming languages.

On DeepSWE v1.1, a long-horizon coding test, it comes out ahead of most larger frontier models. It also beats 3.7 Flash on the Vals Finance Agent V2 and Harvey legal agent benchmarks. On Gray Swan's indirect prompt injection benchmark, Google reports a significant jump in robustness, which for anyone deploying agents may matter more than any leaderboard position.

One caution worth flagging. Several posts published after the announcement circulated numbers like 90.8% on Terminal-Bench 2.1. That benchmark does not appear in Google's announcement at all. A figure repeated across dozens of sites is evidence of copying, not of accuracy, so the numbers above are only the ones in the primary source.

Why the Cyber build is gated

Gemini 3.8 Flash Cyber is not on general sale. Access runs through Google's Fairwind Program, which is limited to trusted government authorities, critical infrastructure operators and software maintainers, and you have to apply for prioritized access.

The reasoning is not subtle. A model that finds real vulnerabilities more than 70% of the time is exactly as valuable to an attacker as to a defender. Google is trying to manage that asymmetry by keeping the offensive-capable build behind an application process while the standard build goes to everyone. It is the same reflex we saw last week when OpenAI shipped Astra with restrictions.

Standard 3.8 Flash, meanwhile, is broadly available: the Gemini API, Google AI Studio, Android Studio, Gemini Enterprise, the Gemini app for Pro and Ultra subscribers, AI Mode in Google Search, and Google Sheets.

Read the calendar, not the launch price

The benchmark table is not the real story here. The calendar is. Announcing in advance exactly when and by how much a model's price will rise is unusual, and to Google's credit it is honest. It also means a team piloting today is looking at a doubled invoice in four months.

The most common costing mistake we see is building the business case on pilot-period invoices. A mid-sized support automation burning 50 million input and 10 million output tokens a month pays roughly $75 today and about $150 from January. On its own that is not painful. At scale, and on a budget somebody already signed off, it becomes a line item you have to defend.

So write the business case at the January price rather than today's. If the numbers work at $1.50 and $7.50, you have a project. If they only work at the introductory rate, what you actually have is a discount.

Sources: Google Blog, Google DeepMind, The Decoder