Gemini 3.8 Flash keeps its price and warns it will spend more tokens

The rate card did not move. Your bill probably will.

Google shipped Gemini 3.8 Flash on 2 September at $0.75 per million input tokens and $3.75 per million output, the same rate as the version it replaces. A second model, Flash Cyber, went to a closed list of defenders through a new Fairwind Program (invitation only, which tells you something).

The scores moved, though. On DeepSWE, the long-horizon software engineering benchmark, it goes to 73.7 percent from 65.3.

The sentence to read twice is in Google's own post

"At times, the model might use more tokens to maximize performance, especially at higher effort levels."

A page earlier, the same post calls that a core design choice. The model "works harder," executing extra reasoning steps and calling tools iteratively. Google then suggests that developers for whom compute efficiency is the primary constraint should turn the effort setting down, or simply stay on the older model.

So a per-token price that holds while per-task consumption rises is a price increase written somewhere the rate card cannot show it. We have no figure for how much more, because Google published none, and that is the gap worth watching.

The cadence is real, and it is fast

Three Flash releases in six weeks. The first landed on 21 July, the second on 13 August, this one on 2 September. Twenty-three days, then twenty.

Across those six weeks the DeepSWE score went from 49.0 to 73.7, a gain of nearly twenty-five points. Simon Willison had the new model wired into his llm-gemini plugin within hours of the announcement, with low, medium and high thinking levels, which is exactly where the token question lives.

Worth reading beside what happened to the Flash models when Artificial Analysis changed its harness a week later.

The cyber model is where Google is more candid than the headline

On CWE-Bench, an external patching benchmark, the cyber model posts a pass rate of 47.2 percent against a leading frontier model's 47.8. Google calls that sitting on the Pareto frontier. True, and also a loss.

The internal numbers read better. A success rate above 70 percent finding vulnerabilities across twenty programming languages. Chrome's security team found it produced 2.6 times more correct patches than the best commercial models, which are a great deal larger than it is. Google's own cloud vulnerability researchers turned up a critical bug in under two hours, on work that usually takes months.

And Wiz, testing on its own penetration-testing benchmark, reported recall up by seven to ten points at two to five times lower cost.

That is the shape of the whole release. Not better than the frontier. Cheaper at nearly the same place, which for most buyers is the better trade.

So how do you price a model that decides how hard to work?

Badly, is our read, at least if you are the one forecasting a budget. How would you even write the line item?

We would expect an independent cost-per-completed-task measurement to show the new Flash costing more than its predecessor on at least one agentic benchmark before the end of October, despite the identical rate. Google publishing median token counts per task would settle it, and it would probably be the single most useful line any lab could add to a model card right now.

And then there is the footnote. That $0.75 is introductory and expires on 31 December. From 1 January it is $1.50 and $7.50.

Double, on a model Google has already told you will spend more tokens than the one it replaces.