Google introduced Gemini 3.6 Flash and the first AI model for cybersecurity. Next up — Gemini 3.5 Pro and Gemini 4

Google has updated its line of language models, introducing several new solutions. The main announcement was Gemini 3.6 Flash, which replaces Gemini 3.5 Flash just a few months after its debut.

At the same time, the company showcased a lightweight version, Flash Lite, and a specialized model for cybersecurity, as well as providing the first insights into the development of Gemini 4.

Gemini 3.5 Flash makes way for the new version

At the Google I/O conference, the main event was Gemini 3.5 Flash, but now it gives way to Gemini 3.6 Flash. The new model becomes the primary choice for developers using the API, as well as for users of the Gemini app.

According to Google, this update was largely driven by community feedback. Following the release of Gemini 3.5 Flash, many developers noted that the model's coding generation capability fell short of expectations.

The likely reason was the company's focus on maximum efficiency and reducing computational costs in light of rising AI usage costs.

Performance has improved, while token consumption has decreased

While the changes cannot be termed revolutionary, Gemini 3.6 Flash demonstrates significant progress across several areas.

In the DeepSWE test, which evaluates programming skills, the new model scored 49% compared to 37% for Gemini 3.5 Flash.

The capabilities for controlling computers have also improved. This feature is now part of the standard capabilities of the Gemini API. In the OSWorld benchmark, which measures the quality of computer operations, the result increased from 78.4% to 83%.

Additionally, Google succeeded in further enhancing the model's efficiency. Despite the improved quality, Gemini 3.6 Flash consumes about 17% fewer tokens than its predecessor.

The changes are particularly noticeable when working with agent scenarios, where the model autonomously performs a sequence of related actions. Google claims that such tasks are now solved more accurately, requiring fewer intermediate steps, and using fewer tokens.

For developers, this means not only more stable application performance but also reduced costs.

The cost of using the API has also decreased:

• incoming tokens — $1.50 per million, unchanged;
• outgoing tokens — $7.50 per million instead of the previous $9.

Gemini 3.5 Flash Lite has become even faster

Despite the release of a new main model, Google continues to develop the 3.5 branch. The company simultaneously introduced Gemini 3.5 Flash Lite — the most economical modern model in its lineup.

Its speed reaches 350 tokens per second, making Flash Lite particularly attractive for large agent systems where high performance and minimal computation costs are crucial.

According to Google, the new Flash Lite's performance now approaches that of flagship models from about a year ago, while remaining significantly cheaper.

The API costs are:

• $0.30 per million incoming tokens;
• $2.50 per million outgoing tokens.

This is slightly higher than the previous Gemini 3.1 Flash Lite, where the corresponding figures were $0.25 and $1.50.

The new version is already available to developers and is gradually appearing in the Gemini app. Additionally, Google plans to actively use Flash Lite in its search engine, where due to its high speed, the model will be able to generate AI Overviews — automatic AI summaries in search results.

Google is preparing the first specialized model for cybersecurity

A separate highlight of the announcements is Gemini 3.5 Flash Cyber — the first Google language model specifically adapted for cybersecurity tasks.

According to the company, it is almost on par with the significantly larger and more expensive model Claude Mythos in terms of finding and fixing vulnerabilities, while maintaining the efficiency characteristic of the Flash family.

At the same time, Google acknowledges that such models fall into the category of dual-use technologies. They can assist not just defenders but also attackers, making it easier to find potential vulnerabilities.

For this reason, there will not be a public release. In the initial phase, Gemini 3.5 Flash Cyber will only be available as part of a limited pilot project within the Google DeepMind CodeMender agent, accessible only to verified company partners and government organizations.

What is happening with Gemini 3.5 Pro

One of the most discussed topics remains the fate of Gemini 3.5 Pro.

During Google I/O, the company stated that the model's release would occur in June, but this has yet to happen. Now Google only mentioned that the new flagship is undergoing testing with a limited number of partners and will be released “as soon as it is ready”.

This model is expected to compete with GPT 5.6, as well as Claude Fable and Claude Sonnet 5.

Earlier media reports indicated that the release of Gemini 3.5 Pro had to be postponed because the model did not show sufficiently strong results in programming tasks compared to its competitors.

Development of Gemini 4 has already begun

However, Google's plans do not end here. The company has confirmed for the first time that it has already begun pre-training Gemini 4.

According to the developers, the new training program will be more extensive than all previous projects in the Gemini family.

When exactly Gemini 4 will be released remains unknown. It is also unclear whether any additional versions in the 3.x series will appear before that. For now, the company only confirms that work on the next generation of models is already underway.