Google moved Gemini 3.8 Flash to general availability on 2 September, its third Flash-tier release in six weeks, alongside a security-tuned sibling called 3.8 Flash Cyber. The launch post promises "significant improvements from 3.7 Flash across software engineering, agentic tasks, and critical, multi-step reasoning in specialized domains." The safety document published beside it says something close to the reverse.
What the model card actually says
The DeepMind model card records that Google's assessments "found that it did not reach any Tracked or Critical Capability Levels," and states verbatim that Gemini 3.8 Flash "does not have meaningful new capabilities or material increases in performance" with respect to the domains set out in the Frontier Safety Framework, compared with Gemini 3.7 Flash. That sentence is not a criticism buried by a reviewer. It is the load-bearing finding: it is precisely because the model does not move on those axes that it clears the framework and ships.
The price is a promotion, not a price
3.8 Flash arrives at $0.75 per million input tokens and $3.75 per million output tokens — the same as 3.7 Flash. The launch post marks that figure as introductory and footnotes the expiry: the rate ends 31 December 2026, and from 1 January 2027 the model costs $1.50 and $7.50. Anyone modelling unit economics off today's page is modelling four months, after which the same tokens cost twice as much.
What the common framing gets wrong
Coverage has read the two documents as if they described different subjects — marketing on one side, compliance boilerplate on the other. They describe the same model, and they are not in conflict once you read what each measures. The launch post is talking about benchmark performance on coding and agentic work. The model card is talking about dangerous-capability thresholds. A model can get materially better at writing software while getting no better at the things a safety framework is built to catch, and that is what Google says happened here. The error is treating "cleared the Frontier Safety Framework" as an endorsement of capability. It is the opposite claim.
Why a third Flash in six weeks
The cadence matters more than the increment. Flash is the tier that carries production traffic, and Google is now revising it faster than most buyers can complete an evaluation cycle. A team that benchmarked 3.7 Flash in July is being asked to re-qualify against a model whose own safety documentation says the capability profile has not meaningfully moved — while the introductory pricing that justified the migration has a hard expiry attached.