Google shipped its fourth Flash model in four months
Gemini 3.8 Flash costs 40% more per task than its predecessor, at the same per-token price, by working longer.
By The Signal · · 5 min read
Gemini 3.8 Flash costs the same per token as the model it replaced, three weeks ago. It costs 40 percent more per task. The difference is not a price change — it is the model doing more work, on its own initiative, and Google billing for the extra effort.
Google released Gemini 3.8 Flash and a specialised variant, Gemini 3.8 Flash Cyber, on September 2, 2026. It is the fourth Flash-tier model in four months, following 3.5 Flash in June, 3.6 Flash in late July, and 3.7 Flash in early August. No new Pro-tier model has shipped since Gemini 3.5 Pro missed its promised June release. Google DeepMind's chief executive, Demis Hassabis, stepped aside to the chairman role in early August; Koray Kavukcuoglu now leads the unit as senior vice president, one rank down from where Hassabis sat.
The model didn't get more expensive. It got more thorough, and thoroughness has a meter running
List price is unchanged: $0.75 per million input tokens, $3.75 per million output tokens, matching Gemini 3.7 Flash's introductory rate. That rate doubles to $1.50 and $7.50 in the new year regardless of which Flash model is running, so the sticker price was never the interesting number here.
The interesting number is what Artificial Analysis, an independent benchmarking firm, measured after running its own evaluation suite: the cost to complete a task rose roughly 40 percent from 3.7 Flash to 3.8 Flash, driven by a 30 percent increase in output tokens per task and more turns on agentic evaluations — meaning the model calls tools, checks results, and tries again more times before returning an answer. Google's own blog post, co-authored by senior product director Tulsee Doshi and Gemini Security Lead Raluca Ada Popa, states the mechanism plainly: "the model might use more tokens to maximize performance, especially at higher effort levels." Developers who want the old cost profile can keep using 3.7 Flash; Google has not withdrawn it.
Think of it as a plumber who now checks the pipe twice before certifying the job. The hourly rate didn't change. The invoice did, because the job took longer.
At $0.58 a task, it undercuts everything scoring near it — for now
Despite the higher per-task cost, Gemini 3.8 Flash remains, by Artificial Analysis's measure, the cheapest model at its intelligence tier: $0.58 per completed task on the firm's Intelligence Index, against $3.76 for Anthropic's Claude Fable 5.1 — a model that scores meaningfully higher (66 versus 59 on the same index, run at high reasoning effort) but at roughly six and a half times the cost per task.
On the index's nine-benchmark composite, Gemini 3.8 Flash's score of 59 puts it level with GPT-5.6 Sol at extra-high effort and Grok 4.6 at medium effort — a three-point gain over its own predecessor. It sits below GLM-5.3 and Kimi K3 (60), Grok 4.6 at high effort (61), GPT-5.6 Sol at max effort (61), Claude Opus 5 (63), and Fable 5.1 (66). Google is not claiming the top score. It is claiming the best score-per-dollar, and on the independent benchmark that claim holds — provided the introductory pricing lasts, which it is not scheduled to.
On the DeepSWE v1.1 long-horizon software engineering benchmark, which measures a model's ability to solve multi-step coding problems without human intervention, Gemini 3.8 Flash scored 73.7 percent, just under Claude Opus 5's 74 percent and ahead of Claude Sonnet 5, GPT-5.6 Sol and GPT-5.6 Terra. It also led on Vals Finance Agent V2 and Harvey's Legal Agent Benchmark — evaluations built around financial analysis and legal-document workflows respectively, the kind of graded, checkable tasks enterprise buyers use to decide which model to route production traffic through.
Fairwind is a membership list, not a product launch
Gemini 3.8 Flash Cyber, the second model released the same day, is tuned specifically for vulnerability hunting and patch generation, and it is not available to the public at any price. Access runs through a new invitation-only arrangement called the Fairwind Program, which Google says has 650 members, including CrowdStrike and the Center for Internet Security. Membership is limited to governments, critical-infrastructure operators, and software maintainers — the institutions Google has decided need automated defense tooling before the general market gets a version of it.
The numbers behind that decision: on the industry-standard benchmark for finding vulnerabilities in C and C++ code, Flash Cyber scored 86.2 percent, ahead of its predecessor and of larger general-purpose models including Anthropic's Mythos 5 and OpenAI's GPT-5.5-Cyber. In real-world vulnerability discovery across more than 20 programming languages, it hit a 71 percent success rate, against 58.9 percent for Gemini 3.7 Flash and 46.6 percent for Gemini 3.5 Flash Cyber. Paired with Google's CodeMender agent, the model reportedly produced 2.6 times more correct patches to vulnerabilities found in Chrome than the best commercial models available, at larger sizes. The company frames CodeMender's job as helping "autonomously find and fix vulnerabilities, protecting critical infrastructure, public services, and national security" — the kind of sentence that reads as marketing until you notice it is restricted to 650 named organizations rather than sold on a price list. That restriction is the actual signal: Google is treating automated exploit-and-patch capability as something to ration, the way export-controlled equipment is rationed, rather than something to ship broadly.
The base model, not the cybersecurity variant, also ships with what Google calls "safeguards against misuse in the domains of Chemical, Biological, Radiological, and Nuclear (CBRN) and cyber offense" — an acknowledgment, buried in a product announcement, that a model good enough to autonomously patch Chrome is also good enough to be asked to do the opposite.
The release cadence is the actual story, and it points at what isn't shipping
Four Flash-tier releases since June, and no Pro-tier model in the same window, is not a coincidence of scheduling. Gemini 3.5 Pro was promised for June and did not arrive. Gemini 3.5 Flash, released instead, drew criticism for benchmark scores that lagged open-weight models from Chinese AI labs. The pattern since has been rapid iteration on the cheaper, faster tier while the flagship tier stays quiet — a strategy that keeps Google visible in weekly benchmark comparisons without requiring the larger training run a genuine frontier model needs.
That strategy has a shelf life. The introductory pricing on Gemini 3.8 Flash doubles in January. Artificial Analysis's cost-per-task advantage assumes the current rate holds; once it doesn't, the "cheapest model at its intelligence tier" claim gets re-measured against whatever Anthropic and OpenAI are charging by then, not against what they charged in September. Nobody has said what happens to the Flash cadence once a real Pro-tier release finally lands — or whether one is coming before the next Flash update makes the question moot again.