On June 17, 2025, Google moved Gemini 2.5 out of preview, making the 2.5 Pro and 2.5 Flash models stable and launching a preview of 2.5 Flash-Lite. The Gemini family is now accessible in Google AI Studio, Vertex AI, the Gemini app, and Search. It retains a one-million-token context while offering finer developer controls.
Flash-Lite is designed for high-volume, latency-sensitive tasks, providing answers in under 100 milliseconds and costing less per token. It surpasses the 2.0 Flash-Lite in coding, math, and multimodal benchmarks. The line maintains its hybrid transformer-retrieval design, with Flash now requiring 20–30% fewer tokens per reply. The 2.5 Pro model introduces Deep Think, a feature that explores multiple hypotheses for complex math or code, along with native audio output, Project Mariner computer-use actions, and enhanced protection against indirect prompt injection.
Introducing the Gemini 2.5 model family:
— Logan Kilpatrick (@OfficialLoganK) June 17, 2025
- Gemini 2.5 Pro (Stable, no changes from 06-05)
- Gemini 2.5 Flash (Stable, updated pricing from 05-20)
- Gemini 2.5 Flash-Lite (Preview, small reasoning model)
More info in 🧵
This rollout follows a spring preview and aligns with Google DeepMind’s tiered strategy that balances speed with capability. Companies like Spline, Rooms, Snap, and SmartBear have run production traffic during testing. Google reports an ELO 1415 on the WebDev Arena coding board and top scores on LMArena preference tests. The stable releases for production applications, along with a budget model and a gated frontier mode, highlight Google’s initiative to offer enterprises and everyday users more options within its AI stack.
