Technology giant Google has launched a major update to its Gemini artificial intelligence family. The company introduced new versions of models aimed at developers and announced a specialized algorithm for cybersecurity tasks. However, the anticipated release of the flagship Gemini 3.5 Pro model has been postponed.

Evolution of the Flash model: efficiency and cost reduction

The previous version, Gemini 3.5 Flash, has been officially deprecated. It has been replaced by the Gemini 3.6 Flash model, which was developed taking into account feedback from the programming community. The main goals of the update were to eliminate errors in coding processes and reduce the cost of generating responses.

Testing results demonstrate significant progress compared to its predecessor:

  • Code automation: In the specialized DeepSWE test, the result increased from 37% to 49%.
  • Interface interaction: The score in the OSWorld test rose from 78.4% to 83%.
  • Energy efficiency: The model requires 17% fewer tokens to perform similar tasks.

Architectural optimization has allowed Google to significantly reduce API usage costs. The new rate is $1.50 per 1 million input tokens and $7.50 per 1 million output tokens. For comparison, the cost of output volume in the previous modification was $9. Currently, Gemini 3.6 Flash is being integrated into the API and is gradually replacing the old version in the web application.

New specialized modifications

Alongside the main version update, the company introduced two new models within the 3.5 lineup.

The first is Gemini 3.5 Flash Lite. This is a high-performance model capable of processing up to 350 tokens per second. It is designed for scaling autonomous agent systems. The API usage cost for this version is $0.30 per 1 million input tokens and $2.50 per 1 million output tokens. The model will also be integrated into the Google search engine to generate AI Overviews blocks.

The second novelty is Gemini 3.5 Flash Cyber. This is the first Google algorithm specifically created for analyzing and protecting digital infrastructure. Developers have combined the high data processing speed of the Flash series with system vulnerability detection functions.

Given the risks of dual use and the possibility of using tools to find vulnerabilities for illegal purposes, Google has restricted public access to the system. The model will be deployed exclusively within a pilot project based on the CodeMender agent from Google DeepMind for a limited circle of partners and government agencies.

Flagship delay and future plans

The launch of the top-tier Gemini 3.5 Pro modification, which is positioned as a direct competitor to GPT-5.6 and Claude Sonnet 5, is being delayed. Currently, the algorithm is in the closed testing phase with a limited group of partners. Exact release dates are not disclosed.

It is likely that the delay is due to the need to refine code generation models to the level of competitors' solutions. Nevertheless, Google has confirmed the start of pre-training for the next generation of systems — Gemini 4. Development involves using a larger-scale computing infrastructure, although the presentation dates for the fourth generation are not yet announced.