Google released Gemini 3.8 Flash on Wednesday, a model that scores 59 on an independent intelligence index while costing a fraction of rival frontier systems, plus a gated cybersecurity variant.
Key Points:
- Gemini 3.8 Flash scores 59 on the Artificial Analysis Intelligence Index, three points above the version shipped three weeks earlier.
- Token prices hold at $0.75 and $3.75 per million until Dec. 31, though cost per task runs about 40% higher than the previous model.
- A separate cybersecurity build reaches only vetted governments, infrastructure operators and software maintainers through a new Fairwind Program.
Gemini 3.8 Flash Benchmark Results
The company introduced both versions on Sept. 2, its third Flash release in six weeks, and both draw on one core system trained heavily on cybersecurity work.
One handles coding and agent tasks, while the other hunts software flaws inside large codebases under far tighter access rules.
Set to high reasoning, the model scores 59 on the Artificial Analysis Intelligence Index, three points above 3.7 Flash and level with GPT-5.6 Sol and Grok 4.6. It also returns 54.9% on HLE-Verified, a test of multi-step reasoning across science, humanities and professional fields. Google points to long-horizon software engineering as the clearest jump.
The company held its introductory rate at $0.75 per million input tokens and $3.75 per million output tokens through Dec. 31. Both of those figures then double on New Year's Day.
Also Read: Claude Fable 5.1 Arrives With $0.25 Cache Reads And Anti-Copy Controls
Why Gemini Costs More Per Task
Cheaper tokens do not translate into cheaper jobs. The model works out to $0.58 per benchmark task, roughly 40% above its predecessor, because average output climbed about 30% to 48,000 tokens and agent runs take extra turns. Rival flagship Anthropic charges close to six times more per task, which keeps the Google system on the cost frontier.
Tulsee Doshi, a senior product director, and Raluca Ada Popa, Gemini security lead at Google DeepMind, said the model simply works harder, adding reasoning steps and calling tools repeatedly. Teams that want the older cost profile can stay on 3.7 Flash, while Pro and Ultra subscribers reach the new model in the Gemini app, in AI Mode and in Google Sheets.
Fairwind Program Gates Cyber Model
The security build sits behind a new Fairwind Program open to vetted governments, infrastructure operators and software maintainers.
On an internal test spanning 20 programming languages it flagged flaws at a success rate above 70%, and Google said its Chrome security team produced 2.6 times more correct patches than far larger commercial rivals. A cloud research team traced a critical vulnerability in under two hours, a job that normally takes months.
The rapid cadence follows a bruising stretch for the AI unit, where a June target for Gemini 3.5 Pro came and went unmet.
Demis Hassabis stepped down as DeepMind chief executive in early August to become chairman, and Koray Kavukcuoglu took charge with the lesser title of senior vice president.
Read Next: Bitcoin ETFs Absorb $216.7M, Undoing A Single Session Of Outflows
