Google introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026. Google positions Flash for software engineering, multi-step reasoning, and longer agent workflows. The Cyber variant focuses on vulnerability discovery and patching, with access restricted to trusted defenders through its Fairwind Program.
The announcement is also explicit about a tradeoff: harder tasks can involve extra reasoning and repeated tool calls, increasing token use. Google lists an introductory Flash price of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, with higher published rates scheduled for January 2027.
Multimodal input expands the workflow
The developer specification lists text, image, video, audio, and PDF input, with text output. It supports up to 1,048,576 input tokens and 65,536 output tokens. Function calling, structured output, and code execution are supported; computer use is listed as a preview. Audio and image generation are not part of this model's output capabilities.
For a design team, that input range suggests useful evaluation tasks: review a recorded onboarding flow, compare screens against a brief, or locate moments of confusion in a product demonstration. These are possibilities to test, not measured results from our own use. Any production workflow still needs representative examples and a review of what the model missed.
Measure the finished task
Our reading is that token price alone is an incomplete way to compare agents. A low-cost call can become expensive if a task repeatedly fails. A more demanding call may be worthwhile if it produces a correct, reviewable result with less intervention. The relevant comparison includes elapsed time, retries, tool execution, and human review.
Teams can make that comparison concrete by choosing a small set of real workflows with clear success conditions. Run the same inputs, preserve the outputs, and record the corrections a person had to make. That produces more actionable evidence than treating a general benchmark score as a guarantee for a specific product.
Long tasks need readable progress
An agent that works through several steps changes what the interface must communicate. Users need to know whether it is reading material, waiting on a tool, revising its approach, or preparing a result. Those states should be visible without exposing an overwhelming stream of internal activity.
A good handoff includes the deliverable, the evidence used, and any remaining uncertainty. For generated code, a successful edit and a successful deployment are separate facts. For a research summary, a retrieved source and a verified conclusion are separate facts. Clear boundaries make a more capable system easier to supervise.
Keep the security variant in context
Flash Cyber's restricted distribution is an important distinction from the general Flash model. The security results in Google's announcement are developer-reported evaluations, not a promise that a codebase becomes secure after an automated pass.
For software teams, AI-assisted findings should enter a reviewable engineering workflow. Reproduction, patch inspection, and regression checks still determine whether a proposed repair is ready. The release creates another option for that work, while the quality of the surrounding process determines how useful the option becomes.