Google's Gemini 3.8 Flash 'Works Harder' -But May Cost More

Google's new Gemini 3.8 Flash keeps 3.7 Flash's per-token pricing but may run up bills by using more tokens on hard tasks. Here's what changed.

Google has released Gemini 3.8 Flash, a new artificial intelligence model that arrived only weeks after its predecessor, Gemini 3.7 Flash. The company says the update "works harder" by taking more reasoning steps on difficult problems and by calling software tools one after another, a design aimed at software engineering and self-directed AI agents.

The per-token price is unchanged from the introductory rate on 3.7 Flash: $0.75 for every million input tokens and $3.75 for every million output tokens. Even so, Google cautions that users could pay more overall, because the model may consume additional tokens to squeeze out better results, particularly when set to higher effort levels. Developers who want to hold token use down can stay on Gemini 3.7 Flash.

That trade-off was picked up quickly by outside observers. Artificial Analysis called 3.8 Flash the least expensive model it has tested at this level of capability, but noted the real cost of a task has climbed roughly 40 percent over 3.7 Flash despite the flat per-token rate. The firm attributed the jump to a 30 percent rise in output tokens per task and a greater number of turns during agent-style tests.

John Ennis, chief executive of Aigora.ai, offered a favourable early take, comparing the model to Anthropic's lineup. He said it delivers "Opus 5 coding quality but at a fraction of the cost and super fast," and predicted it would be especially useful for tasks such as producing Remotion videos.

Benchmark wins over rival models

Google says the model marks significant gains for coding and autonomous agents. It reports that 3.8 Flash beat both its predecessor and other leading models on the DeepSWE v1.1 software engineering benchmark, among them Anthropic's Fable 5, which received an upgrade earlier in the week that promised stronger results at a lower price by reducing the cost of using cached data. Google also says its new model came out ahead of competitors on the Vals Finance Agent V2 benchmark and on Harvey's Legal Agent benchmark.

Cyber edition limited to governments

The model ships with built-in protections against abuse in two areas: chemical, biological, radiological and nuclear threats, and offensive cyber activity. Alongside the general release, Google introduced Gemini 3.8 Flash Cyber and a new Fairwind Program open only to governments and what it calls trusted partners.

The program counts 650 members, including the security firm CrowdStrike and the Center for Internet Security. Members gain access to 3.8 Flash Cyber and to Google's CodeMender agent, which the company says can automatically locate and repair software vulnerabilities to help defend critical infrastructure, public services and national security.

Gemini 3.8 Flash is available now to consumers who hold a Google AI Pro or Ultra subscription, as well as to developers and enterprise customers.

Gemini 3.8 Flash, Google AI, Gemini pricing, AI coding models, DeepSWE benchmark, Fairwind Program, CodeMender, AI agents