Gemini Flash 3.8: Google cuts prices to counter Chinese AIs

Just twenty days separate Gemini 3.7 Flash from its successor. Gemini 3.8 Flash arrived on September 2, accompanied by a version specialized in cybersecurity, bringing the number of Flash models launched in six weeks to three.The August assessment therefore still holds: the fast, low-cost lineup keeps moving up the rankings, while Gemini 3.5 Pro, promised in May, is conspicuous by its absence.

Impressive scores and a bill that quietly balloons

On paper, progress continues. Google claims 73.7% on the DeepSWE coding test, a hair’s breadth behind Claude Opus 5 (74%) and ahead of GPT-5.6 Sol (72.7%), at a price five to six times lower than those two competitors. These figures remain in-house measurements, with no independently published validation, and should be treated as conditional claims.

The pricing, in fact, follows the same pattern as in August: $0.75 per million input text units and $3.75 for output units, a promotion valid through December 31 before doubling on January 1 ($1.50 and $7.50). The back-to-school promotion is becoming a tradition. Except that the model does more work: it uses 30–40% more units per task, raising the actual cost of a typical processing job from $0.40 to $0.58, according to The Decoder. The advertised price goes down; the bill goes up.

The new addition to the lineup is called Flash Cyber, a derivative trained to detect and fix software vulnerabilities, credited with an 86.2% score on the CyberGym test (Claude Mythos had posted 83.1% when it was announced in April). Access is through a program called Fairwind, reserved for governments and “trusted defenders”, whose list Google refuses to publish. Independent researchers and SMEs will have to wait their turn.

Why Google is undercutting prices, and who ultimately pays the bill

Faced with Chinese competition, this pricing strategy remains defensive. DeepSeek charges around $0.14 per million input tokens for its fast version, Qwen can be downloaded free of charge under a permissive license, and GLM, Kimi, and MiniMax sustain a culture of price-cutting that has become industrial strategy. On price alone, Google cannot compete with free.

Its strength lies elsewhere, and that explains the pace: distribution. Gemini 3.8 Flash now powers the search engine’s AI mode, the Gemini app, developer tools, and even Sheets, Google’s in-house spreadsheet. The reasoning is reminiscent of a supermarket’s endcap display: the loss-leader is sold at a loss, while the store makes up the difference on the rest of the shopping basket. An affordable model everywhere means mass adoption; mass adoption means torrents of usage data, the most fiercely contested raw material for training the next generation of models. And for a company that makes its living from advertising targeting, better understanding your queries has never been a losing investment (the terms of use for consumer versions even specify that conversations may be used to improve services).

This industrial pace nevertheless comes at a regulatory cost on the Old Continent: every model marketed in the Union must submit its technical documentation and a summary of its training data under the AI Act. Now, that means a complete file every three weeks.

Until December 31, developers are getting a bargain; in January, their bill doubles. Users of the free versions, meanwhile, will continue paying the usual rate: in data.

👉🏻 Follow the tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: Google