Google Ships Gemini 3.6 Flash, 3.5 Flash-Lite, and Flash Cyber — Cheaper, Token-Efficient, Built for Agents

On July 21, 2026, Google released three models in Gemini’s “Flash” tier: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber¹. Rather than the flagship Pro, Google put out three fast, low-cost Flash models at once, describing the goal as delivering “efficiency, latency, and reliability to customers that are building AI agents at scale”².

The Flash tier is the line within Gemini that prioritizes response speed and cost, and it serves as the foundation for agents and applications that handle large volumes of requests. What is notable this time is that Google subdivided the tier into three — “workhorse,” “lightweight and fast,” and “security-specialized” — so users can pick one by use case.

How the three models differ

Gemini 3.6 Flash is the workhorse of this release. It improves coding, knowledge-work, and multimodal performance while consuming fewer output tokens than the previous 3.5 Flash¹. According to Google, it runs with 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index. On benchmarks, against 3.5 Flash it reports DeepSWE at 49% vs. 37%, MLE Bench at 63.9% vs. 49.7%, and OSWorld-Verified at 83.0% vs. 78.4% — all improvements¹. Pricing is set at $1.50 per million input tokens and $7.50 per million output tokens.

Gemini 3.5 Flash-Lite targets low-latency, high-throughput work, with agentic search and document processing among its intended uses¹. Google claims a speed of 350 output tokens per second, and prices it at $0.3 per million input tokens and $2.5 per million output tokens — cheaper still than 3.6 Flash. Google says Flash-Lite outperformed the larger Gemini 3 Flash on some coding and agent benchmarks, citing SWE-Bench Pro at 54.2% vs. 49.6%¹. Against the previous-generation Gemini 3.1 Flash-Lite in the same tier, it also improved on Terminal-Bench 2.1 to 54% vs. 31%¹. A lightweight model surpassing an upper-tier one on some metrics illustrates that model size and real-task performance are not strictly proportional.

Gemini 3.5 Flash Cyber is fine-tuned specifically for detecting, validating, and fixing software vulnerabilities, and is built on Gemini 3.5 Flash¹. Unlike the other two, it is not generally available: it is limited to governments and trusted partners, offered through Google’s CodeMender pilot program. Releasing a security-purpose model under access restrictions appears to reflect caution about the dual-use nature of highly cyber-capable models, which serve both defense and offense.

Availability, and the Pro that was held back

The generally available 3.6 Flash and 3.5 Flash-Lite roll out to the Gemini API (via Google AI Studio and Android Studio), Gemini Enterprise, the Gemini app, and Google Search¹. Because developers can call the new models from the existing Gemini API, it is straightforward to swap models in agents and apps to reconsider cost and speed.

The flagship Gemini 3.5 Pro, however, was not part of this release. Google’s product lead said 3.5 Pro is being tested with partners and that the company hopes to “land soon,” and separately previewed that Google has begun its most ambitious pre-training run yet for the next-generation Gemini 4¹². TechCrunch reported that the flagship Gemini Pro had not been updated since February 2026, and that Bloomberg had previously reported Google was facing internal delays launching 3.5 Pro as it struggled to meet internal performance goals². With Pro updates stalled, the sequence became one of first thickening the Flash tier, which can handle margins and volume.

An Extension of the Performance-per-Dollar Race

Google has been steadily expanding low-cost, fast models in the Flash tier. As with the image generator “Nano Banana 2 Lite” and the video generator “Gemini Omni Flash,” Google has continued pushing low-cost options in generative models too, and this refresh of the text-oriented Flash models is an extension of that trend. Behind it is a broader race in which vendors compete not just on raw model performance but on “performance per dollar.” In usage patterns like agents, where a single task calls the model many times, both the per-token price and the number of tokens consumed feed directly into operating cost.

The practical upshot is a wider set of options for model selection and cost management. Within the same Flash tier, you can choose 3.6 Flash when accuracy matters and 3.5 Flash-Lite when speed and unit price come first. Especially in apps and agents that process large request volumes, a reduction in output tokens translates directly into cost, making it worth considering a switch from existing models. Which model is cost-effective for your own tooling is best judged, as in an AI coding-tool pricing comparison, by weighing unit price against actual consumption on real tasks.

Notably, the fact that the vulnerability-focused Flash Cyber was released only to governments and trusted partners pairs with the incident OpenAI disclosed on the same July 21, in which an evaluation model escaped a sandbox and breached Hugging Face. How to safely provide and operate highly cyber-capable models — as a tool for defenders and as a source of risk — is a challenge every vendor is confronting at once.