Quick Facts
- Gemini 3.6 Flash reduces output token usage by 17% versus its predecessor, with savings reaching 65% on long-horizon software engineering tasks.
- Google priced the model at $1.50 per million input tokens and $7.50 per million output tokens, down from $9 per million output tokens for Gemini 3.5 Flash.
- Google shipped three models at once: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and the restricted cybersecurity specialist Gemini 3.5 Flash Cyber.
Google released three new Gemini models on July 21, 2026, with no keynote and no claims of a new intelligence frontier. The entire release targets one problem: making AI agents cheaper and more predictable to run in production.
The flagship release, Gemini 3.6 Flash, reduces output token usage by 17% compared to Gemini 3.5 Flash, according to the Artificial Analysis Index, an independent benchmarking group. On DeepSWE, a benchmark measuring multi-step software engineering tasks completed from scratch, token savings reach 65%.
The efficiency gains did not come at the cost of quality. On DeepSWE, Gemini 3.6 Flash scored 49% versus 37% for Gemini 3.5 Flash. On MLE-Bench, it reached 63.9% compared to 49.7%. On OSWorld-Verified, a computer use benchmark, it posted 83.0% versus 78.4%, the highest score in the comparison group, ahead of GPT 5.6 Luna and Grok 4.5.
Long-context performance showed the sharpest gap. On GDM-MRCR v2 at 1 million tokens, Gemini 3.6 Flash scored 54.0% while Gemini 3.5 Flash and Gemini 3.1 Pro did not clear 27%.
Rivals hold ground in specific categories. GPT 5.6 Luna leads on DeepSWE at 67% and Terminal-bench at 84.7%. Grok 4.5 edges ahead on SWE-Bench Pro at 64.7%. Claude Sonnet 5 tops MLE-Bench at 66.9% and GDPVal-AA v2 at 1,607 versus Gemini 3.6 Flash’s 1,421.
Pricing and New Capabilities
Google set the price at $1.50 per million input tokens and $7.50 per million output tokens. The sticker price reduction from $9 per million output tokens is one factor, but the token volume reduction compounds the savings for teams running agents at scale.
Google folded a client-side computer use tool directly into the Gemini API and Gemini Enterprise platforms. Engineers previously built custom intermediary software to let models operate on top of an operating system. That layer is now built in. The model’s knowledge cutoff also advances from January 2025 to March 2026.
Gemini 3.5 Flash-Lite Targets High-Volume Workloads
The second model, Gemini 3.5 Flash-Lite, runs at 350 output tokens per second, the fastest in the 3.5 series by Google’s measure, and is priced at $0.30 per million input tokens and $2.50 per million output tokens. It exposes configurable thinking levels: minimal, low, and higher, letting engineering teams route simple subagent requests to a low-cost path and reserve compute for multi-step work.
On several agentic and coding benchmarks, Flash-Lite outperforms the older Gemini 3 Flash. It scores 54.2% on SWE-Bench Pro versus 49.6%, and 74.0% on OSWorld-Verified versus 65.1%.
A Restricted Model for Cybersecurity
The third model, Gemini 3.5 Flash Cyber, is fine-tuned for finding and fixing code vulnerabilities. It is not publicly available and is designed for controlled security use cases. Google’s CodeMender system runs multiple Flash Cyber sub-agents in parallel to scan large codebases before compiling a single vulnerability report.
Testing on the V8 JavaScript engine found 55 unique confirmed vulnerabilities with Gemini 3.5 Flash Cyber, compared to 47 for Gemini 3.5 Flash and 36 for Opus 4.6. Flash Cyber found 10 issues missed by both competing models.
For software and technology executives evaluating AI agent infrastructure, the release reframes the cost conversation. Token price per million matters less than token volume per completed task, and Google moved both numbers at once with this release.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.