Google Launches Gemini 3.6 Flash and 3.5 Flash-Lite, Flagship Still Delayed

Google announced the launch of two new models in its Gemini family on July 21, 2026: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. It's the company's first major update to its lightweight models since May, though the anticipated flagship Gemini 3.5 Pro still hasn't reached public availability.

Gemini 3.6 Flash is meant to replace the existing 3.5 Flash model as the go-to tool for code, documents, and multimodal tasks. Google says the new model uses 17 percent fewer output tokens than its predecessor while improving answer quality, which lowers the real-world cost of running it at scale.

What the benchmarks show

On DeepSWE, which measures a model's coding skills on real engineering tasks, Gemini 3.6 Flash scored 49 percent versus 37 percent for the previous version. On MLE Bench the score climbed from 49.7 to 63.9 percent, and on OSWorld-Verified, a test of an agent's ability to operate a computer, the model improved from 78.4 to 83 percent. Its GDPval-AA score, which measures the quality of office-style work, rose from 1,349 to 1,421 points.

Gemini 3.5 Flash-Lite is the fastest model in the entire 3.5 family, built for applications that need high throughput and low latency, such as bulk document processing. Google says the model clearly outperforms its predecessor, Gemini 3.1 Flash-Lite, in both coding and agentic tasks, while keeping the lowest price in the lineup.

A third model, for select users only

Alongside the two publicly available models, Google also unveiled Gemini 3.5 Flash Cyber, a specialized model for finding and patching security vulnerabilities, tied to the CodeMender agent. Access is limited to a pilot program for governments and trusted partners, with no public API.

The flagship's shadow

The launch of the new lightweight models coincides with an ongoing delay to Gemini 3.5 Pro. Google had promised the flagship model's release as early as June 2025, but successive rounds of internal testing kept turning up insufficient coding performance. The company reset and updated its training data in late June 2026, but the results disappointed the team once again. Gemini 3.5 Pro is currently available only in testing to select partners and US government institutions, with no public release date given.

The delay has stirred frustration among Google engineers and managers, who worry about losing ground to Anthropic and OpenAI as both push hard on developer tools. Coding has become one of the main battlegrounds among AI labs, and Gemini 3.5 Flash had previously left Google outside the top five strongest models on the Artificial Analysis index.

Gemini 4 teased

Google also said it has already begun what it calls its most ambitious pretraining run to date, for a new model destined for a future Gemini 4 generation. The company gave no timeline or technical details, treating the disclosure as a signal that work on the successor is moving forward independently of the issues with Gemini 3.5 Pro.

The new models rolled out immediately to Google AI Studio, Android Studio, and the Antigravity agentic environment, where Gemini 3.6 Flash appears under the identifier gemini-3.6-flash-tiered. Companies using the Gemini Enterprise Agent Platform and the Gemini app get access to both models, and Gemini 3.5 Flash-Lite is also gradually rolling out to Google Search.

Among the companies praising the new models, Google cites Figma, Harvey, Hebbia and JetBrains for Gemini 3.6 Flash, and Palo Alto Networks and Ramp for Flash-Lite, mainly for their cost-to-performance ratio at scale. For Polish companies and developers using the Gemini API, this translates into lower costs for coding and office tasks, though still without access to Google's strongest model.