Google has launched three new Gemini models, with key updates focused on invocation costs, response speed, and agent-based AI scenarios. However, the much-anticipated Gemini 3.5 Pro has not yet been released, indicating a delay in the product roadmap.
3.5 Pro remains in the testing phase.
At Google's I/O 2026 event in May, the company stated that Gemini 3.5 Pro would be released within a month. However, according to reports, the model failed to meet its targets during internal testing, particularly underperforming on programming tasks, and thus was not launched on schedule.
Bloomberg previously reported that Google attempted to address the issue at the end of June by updating its training data, but the results were unsatisfactory. Google has so far only stated that Gemini 3.5 Pro will be released “when ready.”
3.6 Lower Prices and Improve Efficiency on Flash
The core release of this update is Gemini 3.6 Flash, which continues to be positioned as a lightweight model designed for AI agents and high-frequency use cases. Compared to Gemini 3.5 Flash, this model reduces output token usage by approximately 17% and lowers the output price from $9 to $7.5 per million tokens.
Public benchmark results show that Gemini 3.6 Flash demonstrates significant improvements over its predecessor across multiple tasks, including software engineering, machine learning engineering, and computer operation tasks. Specifically, the DeepSWE v1.1 score increased from 37% to 49%, MLE-Bench rose from 49.7% to 63.9%, and OSWorld-Verified reached 83.0%.
However, in certain programming and knowledge-based tests, OpenAI and Anthropic’s comparable models still lead. The report also noted that Gemini 3.6 Flash still encountered formatting errors and produced results that were not directly usable in simple programming benchmarks.
Flash-Lite and Cyber are designed for different use cases.
Gemini 3.5 Flash-Lite emphasizes high throughput and low cost, making it ideal for document processing, search systems, and batch workflows. This model can output up to 350 tokens per second, with an input price of $0.30 per million tokens and an output price of $2.50 per million tokens.
Another version of Gemini 3.5 Flash Cyber will not be available to the public. Google has restricted access to government agencies and vetted partners for the purpose of identifying and fixing software vulnerabilities.
Gemini 4 has started pre-training.
Google has also confirmed that pre-training for Gemini 4 has begun. Pre-training is the stage in which a model learns foundational data at scale, indicating that Gemini 4 has entered the actual development process, rather than remaining in the planning phase.
Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are now available on the Gemini app, Google AI Studio, and API.