Google ships Gemini 3.8 Flash for coding — the AI model race just got another fast, cheap contender
Google released Gemini 3.8 Flash (codename "Skimaki") on 2 September 2026, a coding- and agent-focused model priced from around $0.75 with a 1M-token context window, positioned to compete directly with Anthropic's Claude Opus 5 and OpenAI's Codex models on software engineering benchmarks — search interest in "Gemini coding" and model comparisons has jumped accordingly.
3 September 2026
Google DeepMind released Gemini 3.8 Flash on 2 September 2026, a model built specifically for coding and agentic workflows, with an introductory price around $0.75 and a 1 million token context window. Early figures put it at 47.2% pass@1 on CWE-Bench, a benchmark for automatically patching known security vulnerabilities, and Google says internal testing shows it producing correct fixes at a notably higher rate than larger commercial models on real-world patch tasks. It arrives roughly six weeks after Gemini 3.7 Flash and follows the now-familiar cadence: a new frontier-adjacent coding model from a major lab every four to eight weeks, each one cheaper or faster or both than what came before.
That cadence is the actual story, more than any single benchmark number. Eighteen months ago, choosing an AI coding assistant was a two-horse race. Now it’s Claude Code, Cursor, GitHub Copilot, Codex, and Gemini’s own tooling, all shipping meaningful updates on overlapping timelines, all claiming benchmark wins that don’t always agree with each other, and all competing partly on capability and increasingly on price per token. For anyone trying to evaluate “which AI coding tool should our team use,” the honest answer keeps changing every few weeks — which is a bad basis for a long-term build decision.
Why the pace of releases is the signal, not the model itself
A model race this fast rewards teams who treat AI coding tools as interchangeable infrastructure — swapped out as better, cheaper options appear — over teams who lock a whole delivery pipeline to one vendor’s roadmap. It also means benchmark claims from any single vendor, Google included, are increasingly a marketing input rather than a reliable predictor of how a model performs on your actual codebase.
So what
If you’re commissioning software, ask your delivery partner how they select and switch between AI coding tools, not just whether they use one. That flexibility is what keeps a project moving when — not if — the next model launch changes the calculus again. See our AI-assisted development approach, or get in touch to talk through your build.