What happened?
On 2 September 2026 Google announced Gemini 3.8 Flash along with its cybersecurity-tuned twin, Gemini 3.8 Flash Cyber. Only three weeks had passed since Gemini 3.7 Flash, making this the third Flash release in six weeks. Over the same period the company's frontier-tier model, Gemini 3.5 Pro, is still nowhere to be seen.
Both variants run on the same underlying model. What separates them is not the architecture but the safety envelope and who is allowed through it: standard Flash is generally available today through the Gemini API, AI Studio, Antigravity, Android Studio and Gemini Enterprise, while Flash Cyber goes only to government agencies, critical infrastructure operators and software maintainers admitted to a new programme called Fairwind.
The token price is flat, the cost per task is not
The most repeated line in the announcement is the price: $0.75 per million input tokens and $3.75 per million output tokens — identical to 3.7 Flash. The introductory rate ends on 31 December 2026, rising to $1.50 and $7.50 from 1 January 2027.
But a flat token price does not mean a flat bill. Google describes the difference as the model "working harder": on complex tasks it runs extra reasoning steps and calls tools iteratively. According to the independent measurement platform Artificial Analysis, output tokens per task rose by 30 percent and cost per task climbed from $0.40 to $0.58, roughly 40 percent. Google's own developer guide says as much outright, recommending that teams stay on 3.7 Flash when compute efficiency is the binding constraint.
- Context window of 1,048,576 tokens and maximum output of 65,536 — unchanged from 3.7 Flash.
- Thinking levels LOW, MEDIUM and HIGH, with MEDIUM the default. MINIMAL has been removed: code that sends it while migrating to 3.8 gets an API validation error.
- Artificial Analysis intelligence index of 59, against 56 for 3.7 Flash.
- At high reasoning the average time per task is 2.5 minutes; 3.7 Flash took 2.2.
What do the benchmarks say?
On DeepSWE v1.1, the long-horizon software engineering benchmark, the table looks like this:
| Model | DeepSWE v1.1 |
|---|---|
| Claude Opus 5 | 74.0% |
| Gemini 3.8 Flash | 73.7% |
| GPT-5.6 Sol | 72.7% |
| Gemini 3.7 Flash | 65.3% |
| Claude Sonnet 5 | 53.8% |
The model also looks resilient to prompt injection: on the Gray Swan IPI benchmark its attack success rate is 5.5 percent. On the same test DeepSeek V4 Pro scores 60.1 percent, Kimi K3 52.7 percent and Grok 4.6 51.8 percent; only Claude Opus 5 edges ahead at 4.8 percent.
Why is the Cyber version locked down?
Flash Cyber scores 86.2 percent on CyberGym, the standard vulnerability discovery benchmark, against 77.5 percent for its predecessor 3.5 Flash Cyber, 83.6 percent for GPT-5.6 Sol and 85.6 percent for GPT-5.5-Cyber. On patching, it reaches 47.2 percent pass@1 on the external CWE-Bench against a leading frontier model's 47.8 percent — Google frames the claim as parity at far lower cost rather than leadership. Chrome's security team reports a 2.6x increase in correct patches, and the Cloud team says it found a critical vulnerability in under two hours, work that normally takes months.
That is also the reason it stays closed: the Cyber variant ships with looser cyber mitigations so it can do defensive work. Google says it prioritised patching over offensive capability, but a model whose safeguards have been relaxed and whose access is restricted cannot be independently audited either — outside Fairwind's 650 members, nobody can test these claims.
What's next?
The only known calendar item is the introductory price ending on 31 December 2026. The open question sits at the frontier tier: while Google shipped three cheap models in six weeks, Gemini 3.5 Pro was reportedly delayed because its coding performance could not match rivals. DeepMind's new head, Koray Kavukcuoglu, says the company is not merely chasing price-performance optimisation and still wants to lead on raw capability.