Google launches Gemini 3.6 Flash, delays 3.5 Pro

Google released three new Gemini Flash models on July 21, 2026, while postponing the previously promised Gemini 3.5 Pro after internal testing found the Pro model did not meet coding performance targets. The new models are available through the Gemini app, Google AI Studio and the company’s API, with the Cyber variant restricted to select government and vetted partners.

The headline release, Gemini 3.6 Flash, is positioned for speed and lower cost in agent and high-volume applications. Google says 3.6 uses 17% fewer output tokens than the earlier 3.5 Flash and reduces output pricing from $9 to $7.50 per million tokens, with input tokens priced at $1.50 per million. Google confirmed the model is live across its consumer and developer channels.

Benchmark results published by Google and third-party testers show mixed outcomes across tasks. On long-horizon software engineering tests, 3.6 Flash scored 49% on DeepSWE v1.1 compared with 37% for 3.5 Flash. On MLE-Bench, it scored 63.9% versus 49.7% for the older Flash. In an environment-control test that measures an AI’s ability to operate a computer to complete real tasks, 3.6 Flash posted an 83.0% score and led the specific test results. Competing models outperformed Gemini on several other benchmarks, including some terminal-based coding evaluations.

Forbes ran a simple coding test on 3.6 Flash that produced an HTML file which failed to render correctly. A small developer team, Deepseek, analyzed the same output, found 11 bugs and implemented eight fixes that produced a playable result. The sequence indicates the model’s high-level approach can be correct while details contain errors that require human correction.

Google also introduced Gemini 3.5 Flash-Lite, a lower-cost, high-throughput variant built for bulk processing. Flash-Lite delivers roughly 350 output tokens per second and is priced at $0.30 per million input tokens and $2.50 per million output tokens. Google described Flash-Lite as aimed at large-scale document workflows and high-volume agent systems. On some coding benchmarks, Flash-Lite outperformed the older 3 Flash-for example, Terminal-Bench 2.1 rose to 54% from the previous model’s 31%.

The third model, Gemini 3.5 Flash Cyber, will not be publicly available. Google restricts that variant to governments and vetted partners that need tools for finding and fixing software vulnerabilities, citing dual-use concerns about broadly releasing capabilities that could be used to exploit code.

Gemini 3.5 Pro, announced last month for imminent release, was held back after internal tests flagged weak performance on coding tasks. A late attempt to adjust training data did not bring the model up to Google’s targets. Google wrote that the Pro tier will ship “as soon as it’s ready.” The company also noted that the delay coincided with a trading session in which Alphabet’s shares fell about 4.4%, reducing the company’s market value by an estimated $200 billion.

Google confirmed that DeepMind has begun pre-training work on Gemini 4. A DeepMind post read: “We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress : )” Google did not provide a firm date for the delayed 3.5 Pro release, reiterating that it will arrive once internal standards are met.