Google splits its new Gemini Flash line by cost, speed and access

Three models answer three deployment questions

Google has added three models to the Gemini Flash family. The interesting part is not that one benchmark number moved past another. It is how clearly the models are being separated by job.

Gemini 3.6 Flash is the general workhorse. Gemini 3.5 Flash-Lite is the cheaper, faster option for large volumes of smaller tasks. Gemini 3.5 Flash Cyber is a specialist for finding and fixing software vulnerabilities, and most developers will not be allowed to use it yet.

This is what the model market increasingly looks like in practice. A company building an agent does not simply choose the smartest model. It chooses where to spend time, money and compute at each step — and which capabilities should sit behind a gate.

The workhorse is selling fewer steps

Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 per million output tokens. Google positions it for coding, document work, multimodal analysis and agents that need to use software tools over several steps.

Its main efficiency claim is not only a cheaper token. Google, citing measurements from Artificial Analysis, says 3.6 Flash produces 17% fewer output tokens than 3.5 Flash on the index's test set. The company also says it takes fewer reasoning steps and tool calls in multi-step work.

That distinction matters. A model can have a low listed price and still become expensive if it writes too much, repeats actions or keeps calling tools. Cost per completed job is more useful than cost per token, though Google has not published a broad, independently audited set of real customer workloads that would settle that measure.

Google reports gains over 3.5 Flash on several tests: 49% against 37% on DeepSWE, 63.9% against 49.7% on MLE-Bench, and 83% against 78.4% on OSWorld-Verified. These are useful signals, not a guarantee that the same gap will appear in a particular product.

Flash-Lite is for the work that arrives by the million

Gemini 3.5 Flash-Lite is priced at $0.30 per million input tokens and $2.50 per million output tokens. It is meant for jobs such as search, translation, classification and document processing, where a small delay or cost repeated across millions of requests becomes a large operational problem.

Artificial Analysis measured it at about 350 output tokens per second, according to Google. The company also reports large improvements over the older 3.1 Flash-Lite on coding, long-context and agent tests.

Both new general models can take text, images, audio and video, and both have context windows of up to one million tokens with text outputs of up to 64,000 tokens. Computer use is built in as a client-side tool, which lets a developer connect the model to an interface and ask it to act there.

Gemini 3.6 Flash and 3.5 Flash-Lite are available now through the Gemini API and Google AI Studio, as well as Google's enterprise agent platform. The workhorse is also in the Gemini app. Flash-Lite is rolling into Google Search.

The cyber model stays behind a gate

Gemini 3.5 Flash Cyber is built on the main 3.5 Flash model and tuned to find, validate and patch software vulnerabilities. Google pairs several instances of it inside CodeMender, an agent system that explores different parts of a codebase and combines the findings into one report.

Google says this repeated, lightweight approach performs competitively with larger models on CyberGym. In a separate test on the V8 JavaScript engine, the company says Flash Cyber found 55 confirmed issues, compared with 47 for the standard 3.5 Flash and 36 for Claude Opus 4.6. Those results come from Google; competitor numbers were self-reported by their providers, and the underlying internal test set is not fully public.

The model is already being used inside Google on Chrome, Android, Cloud, Ads and YouTube, according to the company. Google also says one internal team found remote-code-execution flaws in public APIs and a memory-corruption problem in a production service within two hours. Outsiders cannot verify those examples while the technical details remain private.

Because the same capability can help an attacker, Flash Cyber will enter a limited pilot only for governments and trusted partners through CodeMender. Google says access may expand later. It has not published the admission criteria, price or timetable for that expansion.

What is confirmed, claimed and still open

The releases, prices, access channels and model-card specifications are public. The two general models are usable now. The cyber model is not a general release.

Most quality comparisons come from Google, sometimes using public benchmarks and sometimes internal systems. Artificial Analysis provides an independent speed and efficiency measurement, but no single index captures reliability across long, messy production jobs.

Google's model cards are more restrained than the launch language. They say both general models can hallucinate and may occasionally be slow or time out. Their knowledge cut-off is March 2026. The safety results listed for the models are largely automated internal evaluations, with separate human red-teaming described only at a high level.

The open question is whether the new separation improves real agent economics. Developers will need to measure completed tasks, retries, tool calls, latency and error rates on their own work. A fast model that needs frequent correction may not be cheap. A more expensive one that finishes cleanly may be.

Google has made the choices easier to name. It has not made them automatic.