On July 21, 2026, Google DeepMind released three new Gemini AI models at once — but not the one everyone had been waiting for. We break down exactly what was released, why it matters for developers, and why the flagship Gemini 3.5 Pro still has not appeared.
What Exactly Google Released: Three New Models
All three models belong to the “lighter” Flash class — Google is deliberately developing this tier for fast and inexpensive AI agents, rather than for maximum depth of reasoning.
This is the main model in the lineup, focused on code, knowledge work and multimodal tasks. The main feature is efficiency: the model uses 17% fewer output tokens compared with the previous 3.5 Flash, and on some coding benchmarks the savings reach 65%. The price has also gone down — $1.50 per million input tokens and $7.50 per million output tokens, compared with $9 for output tokens in the previous version.
Gemini 3.5 Flash-Lite — the Fastest and Cheapest
This model is aimed at high-volume tasks under heavy load: 350 output tokens per second, with pricing starting at $0.30 per million input tokens. It is an ideal option for tasks where speed and cost matter more than depth of analysis.
Gemini 3.5 Flash Cyber — a Cybersecurity Model With Limited Access
A specialized model trained to find and fix vulnerabilities in code works inside Google’s CodeMender agent. Unlike the two previous models, access to Flash Cyber is limited — only governments and verified partners within a pilot program, due to the dual-use risks of this technology.
Why Gemini 3.5 Pro Still Has Not Been Released
The most interesting part of this release is not what came out, but what did not. The flagship Gemini 3.5 Pro model, which has been expected for several months, once again did not appear in public access. According to Google’s official wording, the model is in testing with partners, and the company plans to make it broadly available “as soon as it is ready” — without a specific date.
Bloomberg reported internal performance-related delays with the model. Google did not directly deny this information, limiting itself to a general comment about partner testing.
What Comes Next: Google Is Already Starting to Train Gemini 4
In parallel with explanations about the Pro delay, Google reported something unexpected: the company has already begun preliminary training of Gemini 4, calling the process the “most ambitious training cycle” in the history of the lineup. This means that while Gemini 3.5 Pro is stuck at the testing stage, the company’s resources are already partly directed toward the next generation of models.
In Brief: The Key Points
- July 21, 2026 — Google released Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber
- Gemini 3.6 Flash — 17% fewer tokens, cheaper than its predecessor ($1.50/$7.50 per million tokens)
- Flash-Lite — the fastest (350 tokens/sec) and cheapest ($0.30 per million)
- Flash Cyber — a cybersecurity model available only to governments and verified partners
- Gemini 3.5 Pro — still not released, only partner testing, with no exact date
- Google has already begun training Gemini 4 in parallel with the Pro delay
FAQ
When were the new Gemini AI models released? July 21, 2026 — Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber.
Why has Gemini 3.5 Pro still not been released? Google says the model is in partner testing with no specific release date. Bloomberg reported internal performance-related delays with the model.
Who can use Gemini 3.5 Flash Cyber? For now, access is limited only to governments and verified partners through a pilot program — due to the risks of this technology being used for unintended purposes.
How is Gemini 3.6 Flash different from the previous version? It uses 17% fewer output tokens (up to 65% fewer on coding benchmarks) and costs less — $1.50/$7.50 per million tokens versus $9 for output tokens previously.
Has Google started training Gemini 4? Yes, the company officially confirmed the start of preliminary training for Gemini 4, calling it its most ambitious training cycle.
The article was prepared by the TechVisor team — practical IT media for people.
