Google has three new Gemini models on line this time, with an updated focus on call costs, response speed and proxy AI scenarios. However, Gemini, a source of greater outside concern, has still not been published, and the production rhythm has been delayed.

3.5 Pro remains in the testing phase

Google stated at the I/O 2026 conference in May this year that Gemini 3.5 Pro would be launched within a month. It was reported, however, that the model had not met its objectives in internal tests, especially in programming missions, which had not performed as expected, and was therefore not on line on time.

According to Bloomberg, Google tried to fix the problem by updating training data at the end of June, but the results were not ideal. Google is now simply saying that Gemini 3.5 Pro will be released “when ready”.

3.6 Flash lowers prices and increases efficiency

The core version of the release, Gemini 3.6 Flash, continues to be a light model for AI proxy and HF call scenes. Compared to Gemini 3.5 Flash, the output token was reduced by approximately 17 per cent, and the output price fell from $9 per million token to $7.5.

The open benchmarking test showed that Gemini 3.6 Flash had significantly improved over previous generations on several tasks, including software engineering, machine learning engineering and computer operations. Of these, DeepSWE v1.1 scored from 37% to 49%, MLE-Bench from 49.7% to 63.9% and OSWorld-Verified reached 83.0%.

However, OpenAI and Anthropic are still leading in some of the programming and knowledge work tests. It was also mentioned that Gemini 3.6 Flash continued to experience formatting errors in simple programming exercises and that the results could not be used directly.

Flash-Lite and Cyber face different scenes

Gemini 3.5 Flash-Lite places greater emphasis on high-volume throughput and low-cost, suitable for document processing, search systems and bulk workflows. The model can output 350 tokens per second at a price of $0.3 per million token and $2.5 per second.

Another Gemini 3.5 Flash Cyber will not be open to the public. Google limits it to government agencies and audited partners to identify and repair software gaps.

Gemini 4 has initiated pre-training

Google also confirmed that pre-training for Gemini 4 had begun. Pre-training is the stage of large-scale model learning of basic data, which means that Gemini 4 has entered the actual R & D process, not just the planning stage.

Currently, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are online in Gemini applications, Google AI Studio and API.