Google DeepMind released three new AI models on Tuesday, pushing harder on efficiency and specialization while leaving its most anticipated update — Gemini 3.5 Pro — still on the sidelines. The trio of releases includes Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, the company's first model purpose-built for cybersecurity work [1][2].
A Faster, Cheaper Workhorse
The headline release is Gemini 3.6 Flash, which Google describes as its primary "workhorse model" for developers building production applications [2]. It replaces Gemini 3.5 Flash — itself only unveiled at Google I/O in May — which has already been deprecated [1].
Google says the 3.6 revision was shaped directly by user feedback on 3.5 Flash, particularly around code generation, where the earlier model was widely seen as underperforming relative to Google's own marketing [1]. The numbers back up at least some of the improvement: on the DeepSWE coding benchmark, 3.6 Flash scores 49 percent, up from 37 percent for its predecessor [1]. On OSWorld, a test for computer-use capabilities — now a standard feature in the Gemini API — the new model scores 83 percent, compared to 78.4 percent for 3.5 Flash [1].
The more consequential change may be economic. Gemini 3.6 Flash uses approximately 17 percent fewer tokens than 3.5 Flash, a meaningful reduction for businesses running large-scale agentic workflows [1][2]. Output token pricing has dropped from $9.00 to $7.50 per million tokens, while input pricing holds steady at $1.50 per million [1]. Google frames the efficiency gains as benefiting both developers and its own infrastructure costs — a rare alignment of incentives that reflects how seriously the industry has begun scrutinizing the per-token economics of AI deployment [1].
The Ultra-Lean Option
For developers who need to scale agentic systems at minimal cost, Google also released Gemini 3.5 Flash-Lite, which it calls its most efficient modern model [1]. The model hits 350 tokens per second and is priced at $0.30 per million input tokens and $2.50 per million output tokens [1].
Those prices are slightly higher than the previous Flash-Lite generation — $0.25 and $1.50, respectively — but Google argues the capability jump justifies the difference [1]. According to benchmark comparisons, 3.5 Flash-Lite performs roughly on par with what were considered frontier models about a year ago, making it a practical option for high-volume, cost-sensitive deployments [1].
A Cybersecurity Model With Restricted Access
The third release, Gemini 3.5 Flash Cyber, marks Google's first foray into a dedicated cybersecurity AI [1][2]. The model was fine-tuned specifically to identify and remediate security vulnerabilities, and it will not be available to the general public [2]. Instead, Google is running a limited access pilot, restricting the model exclusively to governments and trusted partners [2].
The decision to gate the cybersecurity model reflects the sensitivity of the capability — a tool trained to find exploitable vulnerabilities carries obvious dual-use risks — though Google has not publicly elaborated on the specific criteria for partner eligibility.
The Absence That Overshadows the Announcements
What Google did not release Tuesday may matter as much as what it did. Gemini 3.5 Pro, the company's next flagship model for complex reasoning and coding, was originally promised for June [2]. Google had telegraphed the launch at I/O in May, stating the Pro version was "already being used internally, and we look forward to rolling it out next month" [2].
That rollout never came. Last week, Bloomberg reported that Google was struggling to meet internal performance benchmarks for the model, contributing to the delay [2]. In the meantime, competitors have not been standing still: OpenAI has released GPT-5.5 and begun rolling out GPT-5.6, while Anthropic has launched Claude Opus 4.8, Claude Sonnet 5, and expanded access to its Fable 5 model [2].
Google DeepMind product lead Logan Kilpatrick acknowledged the situation Tuesday, saying the company is currently testing Gemini 3.5 Pro with select partners and hopes to "land soon" [2]. He offered no specific timeline.
Gemini 4 on the Horizon
Kilpatrick did offer one forward-looking signal: Google has begun what he described as its "most ambitious pre-training run yet" for Gemini 4 [2]. No details about architecture, capabilities, or release timing were shared, and the announcement reads more as a confidence signal to the developer community than a concrete roadmap item.
The broader picture Google is painting with Tuesday's releases is one of a company doubling down on the economics of AI deployment — cheaper tokens, faster inference, and specialized models for high-value verticals — while its flagship reasoning model continues to slip. Whether Gemini 3.5 Pro arrives before competitors further widen the capability gap at the top of the market is the question worth watching most closely in the weeks ahead.
