Google Unveils Gemini 3.6 Flash and a Cybersecurity Model, While Gemini 4 Is Already in Training

Google keeps ramping up the pace in the AI-model race. Just a couple of months after unveiling Gemini 3.5 Flash at its May I/O conference, the company has already declared that model obsolete. In its place comes Gemini 3.6 Flash — a bit more capable, better at handling code, and with strong multimodal abilities, according to Google.

The changes in 3.6 are framed as a response to user feedback: the earlier 3.5 Flash failed to deliver on its code-generation promises — likely the price of Google's obsession with efficiency at a time when businesses had started counting the cost of tokens. On the DeepSWE benchmark the new model climbed to 49%, up from its predecessor's 37%, and on the OSWorld computer-control test to 83%, up from 78.4%. At the same time, 3.6 Flash uses roughly 17% fewer tokens, and its output price has been cut: $1.5 per million input tokens and $7.5 per million output tokens, down from the previous $9.

Two more models in the 3.5 line

Alongside the flagship, Google released 3.5 Flash Lite — the most economical of its current models, delivering an impressive 350 tokens per second. In capability it's roughly on par with the cutting-edge models of a year ago, but noticeably cheaper ($0.3 and $2.5 per million tokens). This model, in particular, will see heavy use in AI Overviews search summaries.

Deserving special attention is 3.5 Flash Cyber — Google's first language model tuned specifically for cybersecurity. Google claims it comes close to matching the far larger and more expensive Claude Mythos at finding and fixing vulnerabilities, while retaining the lightness of a Flash model. Acknowledging the “dual-use” nature of such a tool (hunting for holes can be done with malicious intent, too), the company borrows a page from Anthropic and declares the model too dangerous for a public release. It will ship as a limited pilot inside CodeMender, an agent from Google DeepMind available only to vetted partners and governments.

Still waiting on what was promised

As for Gemini 3.5 Pro, which was promised back at I/O for June, it is once again absent from the lineup. Google tersely notes that the flagship meant to compete with GPT‑5.6 and Claude Fable/Sonnet 5 is still being tested by unnamed partners and will ship “when it's ready.” According to early reports, the delay stems from the model falling short of its rivals in programming.

Meanwhile, Google is already looking further ahead: the company has begun pre-training Gemini 4 — a process it describes as more ambitious than anything before it. There's no timeline, nor any clarity on how many more 3.x releases will come out before then.

Source: Ars Technica