Google released Gemini 3.8 Flash on Wednesday, alongside a new cybersecurity-focused variant called Gemini 3.8 Flash Cyber, making this the third Flash model launch in six weeks. The announcement, published on Google's official blog, arrives days after reports that the company was preparing the release, and it comes with benchmark claims that put the budget model within striking distance of rivals that cost several times more per token.
The rapid cadence is the story within the story. Google shipped Gemini 3.7 Flash roughly three weeks ago, and 3.8 Flash now replaces it as the company's default workhorse model. Meanwhile, Google has not released a frontier-level Gemini Pro model since early 2026, and the long-promised Gemini 3.5 Pro remains conspicuously absent. For more context on this story, see our latest AI developments coverage.
Two models, one foundation
Gemini 3.8 Flash comes in two versions. The standard model is positioned as a general-purpose reasoning and coding system for agentic tasks, software development, and high-volume applications. Gemini 3.8 Flash Cyber runs on the same foundation but is tuned specifically for vulnerability detection and mitigation, replacing the earlier 3.5 Flash Cyber.
According to Google's published numbers, Gemini 3.8 Flash scores 73.7 percent on DeepSWE v1.1, a benchmark measuring long-horizon software engineering tasks. That places it just below Claude Opus 5 at 74.0 percent, and ahead of OpenAI's GPT-5.6 Sol at 72.7 percent and Claude Sonnet 5 at 53.8 percent. The previous Gemini 3.7 Flash scored 65.3 percent on the same benchmark, meaning the bulk of this release's gains are concentrated in coding.
Independent benchmarking platform Artificial Analysis gives the new model an Intelligence Index score of 59, three points above its predecessor's 56. That ties it with GPT-5.6 Sol at high reasoning effort and Grok 4.6 at medium reasoning, according to the platform's data, with the gains driven mainly by agentic benchmarks such as tool use and coding.
Aggressive pricing, with a catch
Google is launching the model at an introductory rate of $0.75 per million input tokens and $3.75 per million output tokens, matching the price of 3.7 Flash through the end of the year. From January 2027, the regular price doubles to $1.50 and $7.50 respectively. The comparison with rivals remains stark even then: Claude Opus 5 costs $5.00 per million input tokens and $25.00 per million output tokens, while GPT-5.6 Sol sits at $4.00 and $20.00.
There is a catch worth noting for cost-sensitive teams. Google attributes part of the performance jump to the model running extra reasoning steps on complex tasks and calling tools iteratively — in the company's words, the model "works harder." That translates into higher token consumption. Artificial Analysis pegs the cost per completed task at $0.58, up roughly 40 percent from 3.7 Flash's $0.40, even though the per-token price is unchanged. Google itself recommends developers with tight compute budgets either lower the reasoning level or stay on 3.7 Flash, which remains supported.
Flash Cyber targets security operations
The Cyber variant is the more unusual half of the release. According to Google, internal testing showed Gemini 3.8 Flash Cyber identifying more vulnerabilities and producing working patches more often than its predecessor. Google's Chrome security team reported a 2.6-fold increase in patch accuracy during internal use, and the company says its Cloud team used the model to surface a critical vulnerability within two hours. Google also cited attestations from security partners Wiz and Palo Alto Networks.
Access, however, is restricted. While the standard 3.8 Flash is broadly available starting today, Flash Cyber is limited to trusted testers and governments for now.
The missing frontier model question
The launch puts Google's release strategy under an uncomfortable spotlight. Ars Technica notes that Google has not shipped a frontier-tier Gemini Pro model since early 2026, and reporting earlier this year indicated the release of Gemini 3.5 Pro was delayed after its coding performance failed to match competitors. The Decoder's coverage of the launch framed the rapid Flash cadence as either a sign of engineering momentum or a distraction from the still-missing flagship, depending on who you ask.
Koray Kavukcuoglu, who took over as head of Google DeepMind, pushed back on the framing in the launch materials, stating that Google is not chasing price-performance alone and still intends to lead on raw capability. If the benchmark numbers hold up outside Google's own evaluations, the company may have bought itself time: even its budget Flash tier is now competing with market leaders on coding, the capability enterprises currently care about most.
Not everything improved. On OSWorld-2.0, a benchmark of agentic computer use, 3.8 Flash is better than its predecessor but still trails Claude Opus by a wide margin, according to Google's own charts.
Availability
Gemini 3.8 Flash is available immediately through Google AI Studio, the Antigravity agentic coding tool, and Android Studio for developers, with enterprises getting access via Gemini Enterprise. Consumers can use it in the Gemini app, where it requires a Pro or Ultra subscription, in AI Mode in Google Search, and, for paying subscribers, inside Google Sheets. Free experimentation remains possible through AI Studio.
Google also highlighted the model's improved 3D generation, showing a 3D game reportedly built from a single prompt in Antigravity, with textures generated by the company's Nano Banana image model.
The competitive math is straightforward. Three major labs have now shipped top-tier coding models within weeks of each other, and Google's entry undercuts its rivals on price by a factor of five to seven while claiming comparable coding performance. Whether real-world developer experience matches the benchmarks will determine whether this release quiets criticism of Google's missing flagship or merely postpones it.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
