Google released Gemini 3.8 Flash on September 2, its second Flash model in roughly three weeks and a deliberate push toward cheap, long-running coding work. It arrived with a companion model that almost nobody will be allowed to use.

 

Gemini 3.8 Flash Cyber, a variant tuned to find and patch software vulnerabilities, is distributed only through a new application-gated initiative Google calls the Fairwind Program. The split — one model for everyone, one behind a vetting wall — is the more consequential half of the launch.

 

The general model is live in the Gemini app, AI Mode in Google Search, Google Sheets, Google AI Studio, the Gemini API, Android Studio, Google Antigravity and Stitch. Enterprise customers reach it through Gemini Enterprise on Google Cloud.

 

Related Coverage

 

In Search, availability is narrower. Gemini 3.8 Flash appears as a selectable option in AI Mode for Google AI Pro and Ultra subscribers only, with free-tier users left on the default model.

 

Who Gets Into the Fairwind Program

 

Google is restricting the Cyber variant to applicants it can vet. The program is open to four constituencies:

 

  • Trusted security defenders
  • Government authorities
  • Critical infrastructure operators
  • Software maintainers

 

The model replaces the earlier 3.5 Cyber release for trusted testers. Google says both 3.8 models ship with safeguards against CBRN and offensive cyber misuse.

 

On internal benchmarks spanning 20 programming languages, Google reports the Cyber variant clears 70 percent success at vulnerability detection. It scores 47.2 percent on CWE-Bench patching tasks.

 

Google's own teams supplied the field results. Its Cloud Vulnerability group identified a critical vulnerability in under two hours, and Chrome Security reported the model produced 2.6 times more correct patches to Chrome flaws than what it called the best commercial models that are considerably larger.

 

Coding Claims and the DeepSWE Results

 

The general model's headline claim is about autonomous software engineering rather than raw reasoning. Google points to DeepSWE v1.1, a benchmark measuring end-to-end resolution of engineering tasks.

 

"Outperforms most larger frontier models in autonomously solving complex engineering problems end to end, only at a fraction of the cost."

 

That description comes from Google's launch post announcing the two models.

 

The framing matters because Flash models are not meant to top intelligence leaderboards. They are priced and tuned for volume — the agent loops, code review passes and background tasks where a frontier model's per-token cost becomes the binding constraint.

 

Pricing Before and After January

 

Google is running introductory rates that expire at the end of the year, then doubling them. The published schedule:

 

  • $0.75 per million input tokens through December 31, 2026
  • $3.75 per million output tokens through December 31, 2026
  • $1.50 per million input tokens from January 1, 2027
  • $7.50 per million output tokens from January 1, 2027

 

Independent measurement backs the value argument for now. Artificial Analysis puts the model at roughly $0.58 per Intelligence Index task, making it the cheapest option at its capability level and about six times cheaper than Anthropic's Claude Fable 5.1.

 

Developers building on the introductory rate should model the January step-up before committing to production workloads. A doubling applied to an agentic pipeline running millions of tokens a day is not a rounding error.

 

The Token-Consumption Catch

 

The cost picture is complicated by how the model behaves under load. Google describes the change plainly.

 

"On complex tasks, it exhibits greater diligence — executing extra reasoning steps, and calling tools iteratively."

 

That diligence has a meter attached. The Register reported the model consumes roughly 40 percent more tokens than its predecessor, which erodes part of the per-token savings on exactly the hard tasks Google is marketing it for.

 

Buyers evaluating the model on sticker price will want to benchmark total spend per completed task rather than per million tokens. The two numbers now diverge meaningfully.

 

Ranking Against Claude and GPT-5.6

 

On the Artificial Analysis Intelligence Index, Gemini 3.8 Flash scores 59 in its high-reasoning configuration. That is a three-point gain over 3.7 Flash, which shipped on August 13.

 

The score puts it level with GPT-5.6 Sol and Grok 4.6, and behind Anthropic's Claude Fable 5.1 at 66 and Claude Opus 5 at 63. Google is not claiming the top of the table; it is claiming the best ratio of capability to cost.

 

Google also cites improved robustness against prompt injection on Gray Swan benchmarks, a metric that has become more prominent as agentic deployments expand the attack surface.

 

The release cadence is its own signal. Three Flash iterations in a matter of weeks suggests Google has settled on rapid, cheap model refreshes as its answer to rivals shipping larger and slower.

 

The Fairwind gate is the part worth watching. A model good enough at finding vulnerabilities that its maker will not sell it openly sets a precedent other labs will be asked to match or explain.