Gemini 3.8 Flash is now generally available, with a locked-down cyber sibling for vetted defenders

Google shipped Gemini 3.8 Flash on 2 September 2026, making it the third Flash release in six weeks. It goes straight to general availability under the model ID gemini-3.8-flash, and comes paired with a restricted security variant called Gemini 3.8 Flash Cyber, available only to vetted organisations through Google’s new Fairwind Programme.

Here is what you need to know.

What actually changed from 3.7 Flash

Gemini 3.8 Flash is built on the same architecture as Gemini 3.7 Flash, which launched on 13 August 2026. Google’s own model card points you back to the 3.7 card for training and architecture details. The gains come from harder tuning on agentic behaviour and coding tasks, not a new base model.

The benchmark improvements are meaningful rather than marginal. On Google’s published comparison table:

• DeepSWE v1.1 (software engineering agent tasks): 73.7% vs 65.3% for 3.7 Flash
• Terminal-Bench 2.1 (agentic terminal work): 89.4% vs 85.8%, which also edges past Claude Opus 5 at 89.1% and GPT-5.6 Sol at 88.8%
• OSWorld-2.0 (operating system interaction): 59.0% vs 50.6%
• Vals Finance Agent v2: 61.4% vs 59.0%
• HLE-Verified: 54.9% vs 53.6%

The largest single jump is on BioMysteryBench Human Difficult, up from 43.5% to 56.5%, though that is a narrower research benchmark rather than a practical developer workload.

The honest caveat Google publishes alongside these numbers: 3.8 Flash “works harder” by consuming more thinking tokens to reach its scores. If your priority is cost efficiency over raw capability, Google explicitly recommends staying on 3.7 Flash.

Pricing and the window you need to know about

The introductory price matches Gemini 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens. That rate holds through 31 December 2026. From 1 January 2027, both figures double to $1.50 and $7.50 respectively.

Batch and Flex tiers are available at half the standard rate, and a Priority tier costs 1.8 times the standard rate for guaranteed throughput.

The context window is 1,048,576 tokens with a maximum output of 65,536 tokens. The model accepts text, image, audio, and video input and produces text output. Thinking operates at three levels: low, medium, and high, with medium as the default.

Where you can use it

Gemini 3.8 Flash is available immediately through the generateContent API, Google AI Studio, and the Gemini Enterprise Agent Platform. It has also rolled out to consumer surfaces: the Gemini app for Google AI Pro and Ultra subscribers, AI Mode in Search, and Gemini inside Google Sheets.

What this means for you as a developer

If you are running agentic pipelines or coding assistants, the improvements to DeepSWE and Terminal-Bench scores are the most directly relevant. The model performs well on multi-step terminal tasks and software engineering benchmarks, which matters if you are building anything that involves autonomous code generation, review, or execution.

The pricing clock is the thing worth putting in your calendar. The introductory rate runs to the end of 2026, so if you are planning infrastructure or cost modelling for 2027, build in the doubled rates from 1 January 2027. Catching that late is an avoidable budget problem.

One more thing worth noting: prompt injection resistance has been improved across the 3.8 series. Google evaluated this using the Gray Swan IPI Benchmark. If you are building agents that browse the web or process untrusted documents, that is a practical improvement, not just a marketing note.

The cyber variant and the Fairwind Programme

Gemini 3.8 Flash Cyber is a restricted version of the same model, tuned specifically for cybersecurity work. Access is gated through Google’s new Fairwind Programme, which had over 650 partner organisations at launch. Eligible participants include government agencies and national cyber authorities, critical infrastructure operators in sectors such as healthcare, energy, telecoms, and finance, major software maintainers, and research institutions working on defensive security.

The benchmark numbers are striking. On CyberGym, which tests AI agents against real-world software vulnerabilities, the Cyber variant scored 86.2% against 77.5% for Gemini 3.5 Flash Cyber. Internal testing showed the model producing 2.6 times more accurate vulnerability patches in Chrome than comparable commercial alternatives, and Google’s Cloud Vulnerability Research team used it to find a critical foundational bug in under two hours, a process that previously took months.

Programme participants can use the model for authorised threat simulation, reverse engineering, and defensive malware analysis. They cannot use it for malware creation or redistribution of model access. Organisations must restrict internal access to cybersecurity, incident response, and penetration testing staff.

The model can also be used alongside CodeMender, Google’s vulnerability remediation agent, for end-to-end patch workflows.

This controlled-access approach is becoming standard for frontier security models. Anthropic and OpenAI have both introduced similar vetting requirements for their security-focused model variants. If your organisation works in critical infrastructure or government cybersecurity, the Fairwind Programme is worth investigating directly through Google.

The pace of iteration matters

Three Flash models in six weeks is fast, even by 2026 standards. The shift from multi-month release cycles to near-continuous incremental updates means the model you evaluated two months ago may no longer reflect what is actually in production. Building your evaluation and regression testing into your pipeline rather than treating it as a one-off activity is increasingly the pragmatic approach.

For most developers, 3.8 Flash is a straightforward upgrade on agentic and coding tasks at the same price. For security teams with the right credentials, the Cyber variant represents a meaningful capability step. The main thing to act on before the end of 2026 is your pricing model for next year.