Google DeepMind on July 21, 2026, released three new artificial intelligence models under its Gemini brand: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, the last of which is aimed specifically at finding and fixing software security vulnerabilities. The announcement on the Google blog positions the trio as the company’s answer to developers who need higher token efficiency, lower latency, and more reliable performance when running AI agents at scale.
Gemini 3.6 Flash takes over as the workhorse model
Building on Gemini 3.5 Flash, Google describes 3.6 Flash as its workhorse model, saying it “reduces output token usage by 17% compared to 3.5 Flash” according to the Artificial Analysis Index.
That efficiency gain compounds with a lower sticker price: output dropped from $9.00 to $7.50 per million tokens, with input held steady at $1.50 per million tokens.
On the DeepSWE coding benchmark, the model scores 49 percent, up from 37 percent for 3.5 Flash. Computer use capabilities move from 78.4 percent on OSWorld-Verified to 83.0 percent. The model’s knowledge cutoff advances from January 2025 to March 2026.
Early customers including Hebbia and Harvey cite gains in document parsing, chart and data analysis, and report drafting.
Both 3.6 Flash and 3.5 Flash-Lite feature a one-million-token input context window and a maximum output limit of 64,000 tokens. Both also carry native multimodal input, thinking controls, and built-in Computer Use as a client-side tool via the Gemini API and Gemini Enterprise.
The efficiency story matters especially for the kinds of high-volume agentic workloads that demand throughput over raw reasoning depth, where every extra token multiplies cost across thousands of hourly runs.
Flash-Lite targets speed and high-throughput pipelines
Gemini 3.5 Flash-Lite is designed for high-throughput workloads, delivering 350 output tokens per second according to the Artificial Analysis Index. It is priced at $0.30 per million input tokens and $2.50 per million output tokens at standard API rates.
The model is designed for both low-latency tasks and tasks where high throughput is critical for developer workflows, such as agentic search and document processing, according to Google.
Google’s published numbers show a large step over 3.1 Flash-Lite: 54 percent versus 31 percent on Terminal-Bench 2.1, 72.2 percent versus 60.1 percent on the GDM-MRCR v2 long-context benchmark, and 1,140 versus 642 on GDPval-AA v2.
It significantly outperforms 3.1 Flash-Lite on agentic and coding evaluations and even surpasses the original Gemini 3 Flash on benchmarks including SWE-Bench Pro and OSWorld-Verified.
Flash Cyber brings a security-specialist model under restricted access
Google describes Gemini 3.5 Flash Cyber as a lightweight cybersecurity model built on top of 3.5 Flash and fine-tuned to find, validate, and patch vulnerabilities quickly and efficiently. By building on top of Flash, 3.5 Flash Cyber offers a cost-efficient alternative to large, costly cybersecurity models.
In evaluations, 3.5 Flash Cyber outperforms the mainline Gemini 3.5 Flash and 3.6 Flash when finding new vulnerabilities in codebases. Stress-testing in a Big Sleep evaluation focused on hard-to-find vulnerabilities in complex codebases including Google Chrome and Apple Safari showed it significantly outperformed both Gemini 3.5 Flash and Gemini 3.6 Flash. On the V8 JavaScript engine test, 3.5 Flash Cyber found 55 unique confirmed issues, compared with 47 for Gemini 3.5 Flash and 36 for Claude Opus 4.6.
Within CodeMender, Google’s code security agent, multiple 3.5 Flash Cyber agents work together to produce a single combined report, reaching competitive performance at the frontier on the CyberGym benchmark.
Given the dual-use nature of this technology, Google says it has taken an intentional approach to how it deploys 3.5 Flash Cyber. As part of a limited-access pilot program, the model will be exclusively available to governments and trusted partners via CodeMender. That controlled rollout echoes broader industry discussions about how AI tools with offensive potential, much like advanced GPU-accelerated model infrastructure, require deliberate guardrails before wide release.
Availability and what comes next
Gemini 3.6 Flash and 3.5 Flash-Lite are rolling out via the Gemini app and Google Search. Developers can also access them through Google AI Studio and the Gemini API.
Google first announced 3.5 Pro at its I/O conference in May with the promise of launching it in June. Google says the Pro model is “currently in testing with partners” and that it plans to make it broadly available “as soon as it’s ready.” Google has also signaled it has begun pre-training Gemini 4, describing the new flagship as its workhorse successor for coding, knowledge work, and multimodal tasks.
The three new releases arrive roughly two months after Gemini 3.5 Flash debuted at Google I/O in May 2026, signaling a notably faster cadence of model updates as competition among AI providers intensifies.
