Gemini 3.8 Flash Review of the Facts: Benchmarks, Pricing, Higher Token Use and When to Stay on 3.7
Google's Gemini 3.8 Flash keeps 3.7 Flash's $0.75/$3.75 introductory price but performs more reasoning steps and tool calls, so tasks can cost more. Benchmarks, the Flash Cyber variant and Fairwind Program, cost per task, and practical guidance on effort levels and model choice.
Google released Gemini 3.8 Flash on September 2, 2026, three weeks after Gemini 3.7 Flash and as its third Flash release in six weeks. Google calls it its best reasoning and coding model yet at the same introductory price as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens. The catch is in Google’s own description: 3.8 Flash “works harder,” executing extra reasoning steps and calling tools iteratively, and “might use more tokens to maximize performance, especially at higher effort levels.” The per-token price is the same, but the cost per task can rise. This article summarises the release, the cost implications and how to choose between 3.8 and 3.7.
What Google announced
According to Google’s announcement, the release includes two variants built on the same foundation:
| Gemini 3.8 Flash | Gemini 3.8 Flash Cyber | |
|---|---|---|
| Purpose | General workhorse for coding, agents and reasoning | Vulnerability detection and automated patching |
| Access | Gemini app (AI Pro and Ultra), AI Mode, Sheets, Antigravity, AI Studio, Gemini API | Trusted defenders via the new Fairwind Program |
| Price | $0.75 input / $3.75 output per million tokens (introductory) | Not public |
9to5Google reports that the introductory price runs until December 31, and that the model’s knowledge cutoff is March 2026 for some domains and January 2025 for others.
Benchmarks
Google’s claims for 3.8 Flash:
- DeepSWE v1.1 (long-horizon software engineering): outperforms most larger frontier models in solving complex engineering problems end to end.
- Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark: outperforms 3.7 Flash and other frontier models.
- HLE-Verified: 54.9%, covering multi-step reasoning across STEM, humanities and professional fields.
These are Google-reported results. The Verge noted that Google says the model beat Anthropic’s Fable 5 on DeepSWE v1.1, and that it ships with safeguards against misuse in chemical, biological, radiological, nuclear and cyber offence domains.
Why tasks may cost more
The key point for developers is the difference between price per token and cost per task. The Verge quotes Artificial Analysis describing 3.8 Flash as “the cheapest we’ve measured at this level of intelligence,” while noting that its cost to run the benchmark suite was up about 40% from 3.7 Flash despite unchanged per-token pricing, “driven by a 30% increase in output tokens per task and more turns on agentic evaluations.”
In other words, the model is better value for the capability it delivers, but if your workload does not need the extra capability, you may pay more for the same result. Output tokens are priced five times higher than input tokens, so additional reasoning output has an outsized effect on cost.
Google offers two controls:
- Effort levels. Lower effort levels reduce token overhead.
- Staying on 3.7 Flash. Google says Gemini 3.7 Flash “remains fully supported for efficiency-first workloads.”
When to use 3.8 versus 3.7
| Workload | Suggested choice |
|---|---|
| Long, multi-step coding or agent tasks | 3.8 Flash; the extra diligence is the point |
| Financial, legal or analytical tasks needing careful reasoning | 3.8 Flash, tested on your own cases |
| High-volume classification, extraction or short answers | 3.7 Flash, or 3.8 at low effort |
| Latency-sensitive interactive features | Test both; more reasoning turns can add delay |
| Strict per-request cost limits | 3.7 Flash, or 3.8 with capped effort and monitoring |
A practical approach is to run a sample of real requests through both models and measure tokens per task, total cost, latency and quality. If 3.8 resolves tasks with fewer retries or human corrections, the higher token count may still save money overall.
Gemini 3.8 Flash Cyber and the Fairwind Program
Gemini 3.8 Flash Cyber replaces 3.5 Flash Cyber and is available only to vetted organisations. Google says it shows frontier-level performance on the CyberGym vulnerability discovery benchmark, surpassing significantly larger models, and on an internal benchmark spanning 20 programming languages. 9to5Google reports partner results: Chrome’s security team found it produced 2.6 times more correct vulnerability patches than the best, much larger commercial models; Wiz reported 7.5% to 9.7% higher recall on its penetration testing benchmark at 2.3 to 5.2 times lower cost; and Google’s Cloud Vulnerability Research team found a critical vulnerability in under two hours.
The Verge reports the Fairwind Program is limited to governments and trusted partners, with 650 members including CrowdStrike and the Center for Internet Security, and offers access to Flash Cyber and Google’s CodeMender agent. Restricting the most capable cyber model to vetted defenders mirrors similar approaches by other labs.
Using faster coding models responsibly
More capable, cheaper coding models make it easier to generate large amounts of code quickly. InfoWorld’s guide to vibe coding mistakes cites a survey in which 48% of developers have adopted vibe coding for new projects but 96% do not fully trust that AI-generated code is functionally correct, and a report finding AI pull requests have 1.4 times more critical issues than human ones. Its advice applies to 3.8 Flash too: define requirements before generating code, review and test AI output to the same standard as human code, and watch for security issues.
Practical checklist:
- Set effort levels deliberately per task type rather than using the highest by default.
- Monitor tokens per task after switching models, not just price per token.
- Keep tests and reviews in place for generated code.
- Pin model versions in production so a new release does not change costs unexpectedly.
- Plan for the end of introductory pricing on December 31.
For related reading, see our coverage of the earlier Gemini 3.6 Flash launch, our ChatGPT vs Claude vs Gemini comparison, research on AI coding tool myths, and the AI tools adoption trust gap.
Our analysis
Gemini 3.8 Flash is a capable, low-priced model whose main trade-off is explicit: it spends more tokens to get better results. Google deserves credit for stating that openly and keeping 3.7 Flash available. For teams, the lesson is to budget by cost per completed task rather than price per token, use effort levels to control spending, and reserve the extra diligence for tasks that benefit from it. The rapid release cadence, three Flash versions in six weeks, also argues for pinning versions and re-testing before each upgrade.