Google broadens Gemini range with lower-cost models

Google has launched three lower-cost Gemini artificial intelligence models, including a cybersecurity system designed to detect and repair software vulnerabilities, as the technology group seeks to improve efficiency while its next flagship model remains under development.

The lineup comprises Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber. Google is positioning the releases for organisations that need faster responses, predictable operating costs and the capacity to process large volumes of requests without relying on the most computationally demanding systems.

Gemini 3.5 Flash Cyber is initially being made available to governments and trusted partners through CodeMender, Google’s security-focused coding agent. The restricted rollout reflects the risks attached to models capable of examining software for exploitable weaknesses and generating working attack methods.

The model was fine-tuned to find, validate and patch vulnerabilities. Its comparatively low cost allows security agents to invoke it repeatedly, exploring more code paths before producing a consolidated report. Google sees that approach as an alternative to relying on a single run by a larger and more expensive model.

Testing on Google’s V8 JavaScript engine identified 55 confirmed issues, compared with 47 found by the standard Gemini 3.5 Flash and 36 detected by Anthropic’s Claude Opus 4.6. Ten of the problems discovered by Flash Cyber were not found by either of the other systems tested.

The results highlight an emerging shift in AI-assisted cybersecurity. Rather than concentrating only on maximum capability per request, developers are examining whether smaller models can produce stronger overall results through repeated, specialised searches. The method may make continuous code scanning more practical for large software repositories.

Google has already deployed Flash Cyber within CodeMender to examine internal code used by Chrome, Android, Cloud, Ads and YouTube. During one two-hour exercise, the system detected remote-code-execution weaknesses in public application programming interfaces and a memory-corruption flaw in a production service. It also generated a working exploit capable of bypassing standard memory protections.

Access controls will be central to the model’s wider deployment. A system trained to locate weaknesses can assist defenders, but similar capabilities could be misused to target unpatched infrastructure. Limiting the initial programme to selected public-sector bodies and trusted organisations allows Google to evaluate its safeguards before considering broader availability.

Gemini 3.6 Flash, meanwhile, is being promoted as Google’s general-purpose workhorse for coding, multimodal analysis and agent-based tasks. Independent testing indicates that it can use up to 17% fewer tokens than Gemini 3.5 Flash, reducing the amount of text processed to complete comparable work.

The model is priced at $1.50 per million input tokens and $7.50 per million output tokens. The input rate is unchanged from Gemini 3.5 Flash, while the output price has been reduced from $9. The difference could be material for businesses running customer-service agents, coding tools and document-processing systems at scale.

Gemini 3.5 Flash-Lite targets applications where speed and throughput matter more than advanced reasoning. It costs $0.30 per million input tokens and $2.50 per million output tokens, making it the least expensive model in Google’s 3.5 generation, although its output price is higher than that of the preceding Flash-Lite version.

Both new general-purpose models retain context windows of up to one million tokens, allowing them to process lengthy documents, software repositories and multimedia material. Flash-Lite has been measured at about 350 output tokens per second, supporting workloads that require rapid generation across thousands of simultaneous requests.

The launches arrive one day before Alphabet is due to report quarterly earnings, placing attention on the commercial returns from its heavy investment in data centres, custom processors and AI development. Alphabet has raised its 2026 capital spending guidance to between $180 billion and $190 billion as demand for cloud computing and AI infrastructure expands.

Investors are also watching the delayed Gemini 3.5 Pro, which had been expected in June. The flagship model remains in testing after falling short of internal expectations on coding tasks, an increasingly important measure for enterprise customers. Google has not provided a new release date, saying only that development continues.