Google has released Gemini 3.8 Flash, a new general-purpose model aimed at developers and organisations that need capable AI agents without the latency and cost usually associated with larger frontier systems. Announced on 2 September 2026, the model focuses on long-running software tasks, tool use, multimodal analysis and complex knowledge work.
The launch also includes Gemini 3.8 Flash Cyber, a more specialised version for finding and patching software vulnerabilities. That edition is not broadly available: Google is limiting it to trusted defenders through its new Fairwind Program.
Gemini 3.8 Flash: the background
“Flash” models are designed to balance intelligence, speed and operating cost. That balance matters for production systems because an autonomous agent may make repeated model calls, inspect files, use tools and revise its work before completing one request. A modest difference in token use or latency can become significant at scale.
Gemini 3.8 Flash builds on Gemini 3.7 Flash, which arrived only weeks earlier. Google describes the new release as its strongest workhorse model yet for coding and agents, rather than a replacement for every Gemini tier. The previous model remains supported for workloads where efficiency is more important than maximum reasoning quality.
What changed in Gemini 3.8 Flash?
Stronger long-horizon coding and agent workflows
The main upgrade is sustained performance on tasks that require many connected steps. Google says Gemini 3.8 Flash improves software engineering, agentic work and specialised reasoning. It can call tools iteratively and spend more reasoning effort when a problem is difficult.
On Google’s published evaluations, the model scored 54.9% on HLE-Verified, a demanding multi-domain reasoning benchmark. Google also reports gains on DeepSWE v1.1 for long-horizon software engineering, plus finance and legal agent benchmarks. These are useful indicators, but buyers should still test the model on their own code, documents and workflows rather than treating vendor benchmarks as universal results.
Multimodal input and a large context window
The official model card lists support for text, images, audio, video and PDF inputs. Gemini 3.8 Flash accepts up to one million input tokens and can produce up to 64,000 output tokens. That makes it suitable for reviewing large codebases, lengthy reports, meeting media or collections of business documents in one workflow.
It also supports function calling, search as a tool and computer use. Those capabilities are important for agents that must do more than generate text—for example, checking a data source, updating an internal application or navigating a controlled interface.
New specialist cyber-defence model
Gemini 3.8 Flash Cyber concentrates on vulnerability discovery and automated patching. Google reports a 47.2% pass@1 score on the external CWE-Bench patching benchmark. In Chrome security testing, the company says the model produced 2.6 times more correct vulnerability patches than the best larger commercial models it compared against.
Because a powerful cyber model can be misused, Google is providing the specialist edition through Fairwind. The limited-access program prioritises government authorities, critical-infrastructure operators and software maintainers, with operational controls such as restricted staff access and multi-factor authentication.
Gemini 3.8 Flash pricing and availability
Gemini 3.8 Flash is generally available through the Gemini API and Google AI Studio, as well as Google Antigravity, Android Studio and Gemini Enterprise. Consumers can access it through eligible Google AI Pro and Ultra subscriptions in the Gemini app, AI Mode in Search and Gemini in Google Sheets.
Google’s introductory API price is US$0.75 per million input tokens and US$3.75 per million output tokens—the same introductory rate as Gemini 3.7 Flash. The offer runs through 31 December 2026. Google says rates will increase on 1 January 2027 to US$1.50 per million input tokens and US$7.50 per million output tokens.
Teams planning a long-term deployment should therefore model both the introductory and scheduled prices. They should also measure total task cost, not simply price per token, because Google notes that the model may use additional reasoning steps and tokens at higher effort settings.
Why the release matters
The significant part of this launch is not just another benchmark increase. It is the combination of stronger reasoning with a tier intended for high-volume use. Many useful AI applications require an agent to remain reliable across dozens of actions. A cheap first response is less valuable if the system fails halfway through a coding, research or service workflow.
For developers, Gemini 3.8 Flash could support code migration, test generation, bug investigation, interface prototyping and repository-wide analysis. Businesses may use it for document-heavy operations, internal support, reporting or controlled workflow automation. Security teams outside Fairwind can still use publicly available Gemini models with CodeMender, according to Google, although the specialist Cyber model remains restricted.
Risks and limitations
Gemini 3.8 Flash is still a foundation model and can hallucinate. Its large context window does not guarantee that every detail in a long input will be interpreted correctly, and agents with tool access can amplify mistakes by acting on them. High-impact operations should use approval gates, narrow permissions, audit logs and rollback plans.
The model card also notes possible slowness or timeouts and says multilingual safety performance regressed slightly against Gemini 3.7 Flash in an automated evaluation. Its stated knowledge cutoff is March 2026 for some domains, while knowledge in others may be limited to January 2025. Search or another verified data source is therefore necessary for current facts.
Cybersecurity claims deserve particular care. Google’s results include internal testing, and real-world performance will vary by language, repository and vulnerability class. Automated patches still need review, testing and staged deployment.
What to watch next
The next questions are how the model performs in independent tests, how much token usage rises at higher effort levels and whether the scheduled 2027 price affects adoption. It will also be worth watching how quickly Google expands Fairwind access and whether restricted cyber models measurably reduce the time between discovering and repairing serious flaws.
Conclusion
Gemini 3.8 Flash is a practical release for organisations building AI agents at scale. It combines a one-million-token context window, multimodal inputs, tool use and improved coding performance with relatively low introductory pricing. The specialist Cyber edition also shows where frontier capabilities may increasingly arrive through controlled-access programs rather than unrestricted APIs.
The strongest adoption strategy is straightforward: benchmark it against real tasks, calculate full workflow costs, and keep human oversight around consequential actions. Better agent performance is valuable, but dependable deployment still depends on permissions, evaluation and operational discipline.