Google Rapidly Iterates Gemini with 3.6 Flash Family and Restricted Cybersecurity Variant

On July 21, 2026, Google released Gemini 3.6 Flash and 3.5 Flash-Lite, simultaneously deprecating 3.5 Flash — the model it had featured at Google I/O just two months earlier [2][1]. A restricted cybersecurity variant, 3.5 Flash Cyber, was announced July 17 and limited to governments and trusted partners because of its demonstrated ability to autonomously find and exploit production vulnerabilities [3]. Gemini 3.5 Pro, expected in June, remains unreleased; Bloomberg reported the delay stems from coding performance falling short of internal targets, and Alphabet shares fell on the news [4][5]. Google also disclosed it has begun pre-training for Gemini 4, described as its most ambitious run yet [1].

Google is cycling through its Flash models faster than typical product timelines suggest — deprecating a flagship within two months of its announcement — while a quality shortfall holds its Pro-tier model off the market. The cybersecurity model introduces a new category of constrained-access AI capable of generating working exploits autonomously; how Google manages that access, and whether the controls hold, is a live question for the broader industry.

Google released three new Gemini models on July 21, 2026, reorganizing its Flash tier and introducing a restricted cybersecurity variant. Gemini 3.6 Flash replaces 3.5 Flash — which had been the centerpiece of Google I/O just two months earlier — claiming 17% fewer output tokens while improving on the benchmark areas where 3.5 Flash underperformed, including DeepSWE coding (49% vs. 37%) and OSWorld computer use (83.0% vs. 78.4%) [1]. Gemini 3.5 Flash-Lite, released the same day, targets high-volume latency-sensitive workloads, reaching 350 output tokens per second while outperforming the older Gemini 3 Flash on SWE-Bench Pro (54.2% vs. 49.6%) [1]. Ars Technica noted the unusual speed of 3.5 Flash's deprecation and suggested Google's intense focus on token efficiency for enterprise cost reasons may have contributed to the original coding shortfall that 3.6 Flash was built to address [2].

Four days earlier, Google had announced Gemini 3.5 Flash Cyber, a model trained specifically for offensive and defensive security work. In internal evaluations, it found 55 unique confirmed vulnerabilities in Chrome's V8 JavaScript engine, compared to 47 by mainline 3.5 Flash and 36 by Claude Opus 4.6, across a fixed number of invocations [3]. In a two-hour evaluation on Google's own production systems, it uncovered remote code execution vulnerabilities in public APIs and a memory-corruption flaw in a sensitive service, then generated a functional RCE exploit bypassing both ASLR and W^X mitigations [3]. Google is limiting the model to governments and trusted partners through its CodeMender platform, framing access controls rather than capability reduction as the appropriate safeguard for dual-use risk [1][3].

Running alongside these releases is the unresolved delay of Gemini 3.5 Pro, which had been expected to launch in June. Bloomberg reported that Google held the model back because coding performance fell short of internal targets, and Alphabet's share price fell on that report [4][5]. Google's announcement briefly addressed the Pro situation without specifics, and separately disclosed that the company has begun what it called its most ambitious pre-training run yet for Gemini 4, with no timeline or capability claims attached [1].