Gemini 3.8 Flash Cyber and the Fairwind Program: Google Joins OpenAI and Anthropic in the Same Ten-Day Window

On September 2, Google released Gemini 3.8 Flash, its third Flash version in six weeks, alongside a locked-down security variant called Gemini 3.8 Flash Cyber. The model beats its predecessor on every benchmark Google published and beats Claude Opus 5 on three of them. More interesting than the numbers themselves is a program called Fairwind, through which Google gives access to the model's strongest cybersecurity capabilities only to vetted partners, in the exact same ten-day window in which OpenAI and Anthropic built identical solutions for their own models.

Numbers that actually add up

Gemini 3.8 Flash offers an input context of up to 1,048,576 tokens, a 65,536-token output limit, multimodal support, and configurable "thinking" levels: low, medium, and high. On Terminal-Bench 2.1 the model scores 90.8%, against 81.6% for the previous 3.7 Flash version. On DeepSWE v1.1, a benchmark for long-horizon coding tasks, the score is 73.7% against 65.3%. On OSWorld-2.0, a computer-use test, it's 59% against 50.6%. These are real, measurable jumps, not the cosmetic tweaks you'd expect from a third update to the same model line in under six weeks. Beating Opus 5, a significantly more expensive, larger Anthropic model, on three benchmarks with a model from the Flash family, theoretically Google's cheaper, faster product line, shows that the split between "flagship" and "fast" models is stopping being an accurate map of actual capability.

A promotional price set to expire with the new year

Gemini 3.8 Flash currently costs $0.75 per million input tokens and $3.75 per million output tokens, but only through December 31, 2026. On January 1, 2027, both rates double, to $1.50 and $7.50. Google isn't hiding this in fine print, the promotional pricing is labeled as temporary outright. It's an aggressive market-entry tactic: a low price up front so developers build their products around the model before the cost of running it jumps 100%.

That same clock has been ticking since July. Gemini 3.6 Flash launched July 21, Gemini 3.7 Flash followed three weeks later on August 13, and wasn't a new model built from scratch at all, just an algorithmic improvement on the same 3.6 reasoning foundation. Both of those versions carried exactly the same promotional pricing and exactly the same expiration date, December 31, 2026. Gemini 3.8 Flash isn't setting a new clock, it's holding to a deadline set two months earlier, even though the model itself has now changed three times in the meantime.

Fairwind: the same idea as Daybreak and the Cyber Verification Program, different name

Gemini 3.8 Flash Cyber, available exclusively through the Fairwind Program, pairs the model with CodeMender, Google DeepMind's own agent for automated vulnerability patching, which has been running since July and already has 72 real security fixes shipped into open-source projects to its name. CodeMender doesn't stop at detecting a bug: it builds a working proof-of-concept exploit in a controlled sandbox to filter out false positives, generates a fix, and a separate critique sub-agent checks that fix for new regressions before it ever reaches a human for final approval. The Cyber version of this same model hits more than 70% success on real-world vulnerability discovery and sits on the Pareto frontier of the CWE-Bench benchmark, which measures the quality of automated code patches. Access requires a formal application, background checks, and acceptance of the program's operating rules. At launch, Fairwind covers government agencies, Google Cloud customers, and cybersecurity partners, with priority for critical infrastructure operators and maintainers of widely used software.

OpenAI built Daybreak Access for GPT-6 Astra a day earlier, and Anthropic built its Cyber Verification Program for Mythos 5.1 a day before that. Three different companies, days apart, independently arrived at exactly the same conclusion: the strongest offensive and defensive cyber capabilities of new models need to sit behind a formal, vetted wall instead of being available to anyone with an API key. None of the three copied the solution from a competitor, since all three programs were built in parallel, which on its own is stronger evidence that this is a genuine, shared industry response rather than one company's marketing trend.

A pace that became its own story

Gemini 3.8 Flash is the third update to the Flash line in six weeks, after 3.6 and 3.7. That release frequency would look like chaos, not progress, in any other tech industry. In AI it has become normal enough to pass almost without comment, even though this exact pace is what Jacob Coxon pointed to in his viral resignation warning from Anthropic as proof that no company has time left for a calm risk assessment before the next launch.