Google’s Gemini AI Models Cut Costs 17% to Undercut GPT-5.6

Google has expanded its Gemini AI models lineup with three new releases — Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — each targeting a distinct gap in the market: efficiency at scale, raw speed, and a closely guarded play in AI-powered cybersecurity. The announcements, made on July 21, 2026, arrived on the eve of Alphabet’s earnings report and against a backdrop of intensifying pressure from Chinese rivals and Anthropic’s growing lead in automated code defense.

Key takeaways

• Gemini 3.6 Flash reduces output token usage by 17% compared to 3.5 Flash, priced at $1.50 per million input tokens and $7.50 per million output tokens.
• Gemini 3.5 Flash-Lite is the fastest model in the 3.5 family, delivering 350 output tokens per second, priced at $0.30 per million input tokens and $2.50 per million output tokens.
• Gemini 3.5 Flash Cyber is fine-tuned for cybersecurity vulnerability detection and remediation, available exclusively to governments and trusted partners via a limited pilot through CodeMender.
• Both 3.6 Flash and 3.5 Flash-Lite are live today across the Gemini API, Google AI Studio, Android Studio, Gemini Enterprise Agent Platform, the Gemini app, and Google Search.
• Gemini 3.5 Pro is in partner testing, and Gemini 4 pre-training — described as Google’s most ambitious yet — has already begun.

Google launches three new Gemini AI models

The trio of releases represents a calculated push across different parts of the AI market simultaneously. Rather than one headline flagship, Google is shipping models tuned for specific workloads — a strategy aimed at capturing developers running high-volume production traffic, enterprises building autonomous agent pipelines, and government cybersecurity teams defending critical infrastructure.

Gemini 3.6 Flash: more output for less cost

Gemini 3.6 Flash is the centerpiece of the launch. Building directly on feedback from 3.5 Flash users, the new model improves coding accuracy, multimodal performance, and knowledge work — while simultaneously becoming cheaper to run. According to the Artificial Analysis Index, 3.6 Flash consumes 17% fewer output tokens than its predecessor. On some coding benchmarks, specifically DeepSWE by Datacurve, the reduction reaches up to 65%.

The practical upside for developers is straightforward: less token usage means lower per-task costs. At $1.50 per million input tokens and $7.50 per million output tokens, 3.6 Flash comes in cheaper than 3.5 Flash, and according to Google’s figures, also undercuts comparable models on cost per task, as measured by Artificial Analysis data.

Performance gains are visible across several benchmarks. On DeepSWE, 3.6 Flash scores 49% versus 3.5 Flash’s 37%, reflecting fewer unwanted code edits and tighter execution loops. In machine learning research tasks measured by MLE Bench, it jumps to 63.9% from 49.7%. Computer use — now a built-in client-side tool via the Gemini API and Gemini Enterprise — also improves, reaching 83.0% on OSWorld-Verified compared to 78.4% previously. Customers including Hebbia and Harvey have flagged its particular strength in multimodal tasks such as document parsing, chart analysis, and report drafting.

Safety safeguards built into 3.6 Flash

The model ships with enhanced Frontier Safety safeguards covering chemical, biological, radiological, nuclear, and cyber offense domains. These protections are designed to make 3.6 Flash substantially more resistant to jailbreaks, while still minimizing refusals for legitimate uses — a balance that matters for enterprise customers deploying the model in sensitive or regulated environments.

3.5 Flash-Lite: speed at scale

For developers where throughput is the priority over raw intelligence, Gemini 3.5 Flash-Lite fills the gap. It is the fastest model in the 3.5 series, running at 350 output tokens per second as measured by Artificial Analysis — making it suited for high-volume tasks like agentic search, document processing, and subagent coordination within larger AI pipelines.

Priced at $0.30 per million input tokens and $2.50 per million output tokens, it delivers significantly better quality than the prior 3.1 Flash-Lite generation. On Terminal-Bench 2.1, it scores 54% versus 31% for its predecessor, and on GDM-MRCR v2 for long-context tasks, it reaches 72.2% compared to 60.1%. Notably, it also outperforms the more expensive 3 Flash on several agentic evals — including SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74.0% vs. 65.1%) — giving developers running 2.5 or 3 Flash workloads a meaningful reason to consider switching down the model tier for cost savings without a quality penalty.

Like 3.6 Flash, it includes computer use as a built-in tool, enabling it to handle agentic tasks reliably across surfaces. Developers can also configure its thinking level: minimal and low thinking settings prioritize speed and low cost for bulk tasks, while higher thinking levels handle multi-step subagent workloads that require more deliberate reasoning.

Specialized cybersecurity model and restricted access

Gemini 3.5 Flash Cyber and the CodeMender agent

The most strategically pointed release is Gemini 3.5 Flash Cyber, a model built on the 3.5 Flash base and fine-tuned specifically for cybersecurity vulnerability detection and remediation. It works inside CodeMender, Google’s code security agent, where multiple 3.5 Flash Cyber agents collaborate to produce a single consolidated vulnerability report. The system delivers competitive performance on the CyberGym benchmark at a lower price per token than larger generalist models.

The timing is pointed. As reported by CNBC, Anthropic has built an early lead in automated code defense with its Mythos model, and Google’s new cybersecurity-focused release is its clearest challenge to that position yet. AI models have become capable of finding security vulnerabilities faster than human teams can patch them, which is precisely the problem 3.5 Flash Cyber is designed to address at scale.

Access limited to governments and trusted partners

Because of the dual-use nature of a model trained to find software vulnerabilities, access to 3.5 Flash Cyber is deliberately restricted. It will be available exclusively through CodeMender to governments and trusted partners as part of a limited-access pilot program. The rationale is giving frontline defenders a head start on identifying and patching critical vulnerabilities before bad actors can exploit them, while limiting broader exposure to potential misuse.

This controlled rollout also signals how Google is thinking about deploying sensitive AI capabilities more broadly — not as open APIs, but through curated partnerships with verifiable accountability.

Where the models are available now

Gemini 3.6 Flash and 3.5 Flash-Lite are available immediately across multiple channels:

• Developers can access both via the Gemini API, Google AI Studio, and Android Studio.
• Enterprises can reach both through the Gemini Enterprise Agent Platform.
• General users can access them through the Gemini app, with 3.5 Flash-Lite rolling out in Google Search.

The broad distribution across developer tools, enterprise platforms, and consumer products reflects how seriously Google is treating the Flash tier — not as a stripped-down alternative, but as the primary workhorse for scaled AI deployment.

What comes next: Gemini 3.5 Pro and Gemini 4

Google has also offered a rare degree of roadmap transparency alongside today’s launches. Gemini 3.5 Pro is currently in testing with partners, with broader availability planned once ready. More significantly, the company confirmed it has begun its most ambitious pre-training run yet for Gemini 4 — described internally as the next generation of models, with Google expressing confidence in early progress.

The broader context matters here. Chinese rivals are moving fast: Moonshot AI’s Kimi K3 recently generated enough demand to force subscription and API limits, while Alibaba is advancing its Qwen model line. Google’s response is to compete on cost efficiency and model breadth rather than waiting for a single flagship. That strategy — shipping capable, cheap, widely distributed models while the most ambitious work continues in the background — is increasingly becoming the company’s defining competitive posture in AI. Whether that posture holds as Gemini 4 comes online, and whether 3.5 Pro arrives before rivals close the gap, will define how this moment is remembered.

FAQ

What are the key improvements of Gemini 3.6 Flash compared to 3.5 Flash?

Gemini 3.6 Flash reduces output token usage by 17%, delivers better coding and knowledge work performance, and lowers cost per token, priced at $1.50 per million input tokens and $7.50 per million output tokens — cheaper than 3.5 Flash.

Who can access the Gemini 3.5 Flash Cyber model?

Access to Gemini 3.5 Flash Cyber is restricted exclusively to governments and trusted partners through a limited pilot program delivered via CodeMender.

Where are Gemini 3.6 Flash and 3.5 Flash-Lite available?

Both are available via the Gemini API, Google AI Studio, Android Studio, the Gemini Enterprise Agent Platform, the Gemini app, and Google Search.

What safety features does Gemini 3.6 Flash include?

It includes enhanced Frontier Safety safeguards targeting chemical, biological, radiological, nuclear, and cyber offense misuse, designed to reduce jailbreak risks while minimizing refusals for legitimate, beneficial uses.

Article produced with the assistance of artificial intelligence and reviewed by the editorial team.