Abstract:

On September 25, 2025, Google announced the release of updated preview versions of its Gemini 2.5 Flash and Flash-Lite models, available on Google AI Studio and Vertex AI. The updates focus on improving model quality and efficiency. Gemini 2.5 Flash-Lite has been enhanced for better instruction following, reduced verbosity (leading to a 50% reduction in output tokens), and stronger multimodal and translation capabilities. The updated Gemini 2.5 Flash shows significant improvements in agentic tool use, reflected in a 5% gain on the SWE-Bench Verified benchmark, and is more cost-efficient, using 24% fewer tokens. To simplify access for developers, Google introduced a "-latest" alias for model families, which automatically points to the most recent preview version. The company clarifies that these are preview releases intended for testing and feedback to shape future stable models, while the current stable versions remain recommended for production use.

Google Announces Key Updates to Gemini 2.5 Flash and Flash-Lite Models

• September 25, 2025: New Preview Models Released: Google has released updated preview versions of Gemini 2.5 Flash and Gemini 2.5 Flash-Lite on Google AI Studio and Vertex AI, aimed at improving quality and efficiency.
• Gemini 2.5 Flash-Lite Improvements: This version is significantly better at following complex instructions, produces more concise answers (reducing output tokens and cost by 50%), and features more accurate audio transcription, image understanding, and translation.
  • Model String: gemini-2.5-flash-lite-preview-09-2025
• Gemini 2.5 Flash Enhancements: This model update focuses on two key areas:
  • Better Agentic Tool Use: It shows improved performance in complex, multi-step applications, with a notable 5% performance gain on the SWE-Bench Verified benchmark (from 48.9% to 54%).
  • Increased Efficiency: The model is more cost-efficient, achieving higher quality with 24% fewer tokens, thus reducing latency and cost.
  • Model String: gemini-2.5-flash-preview-09-2025
• Simplified Access with "-latest" Alias: Google introduced a new -latest alias (e.g., gemini-flash-latest) that always points to the most recent model version. This allows developers to experiment with the latest features without needing to update their code for each preview release.
• Preview vs. Stable Models: These releases are previews intended for developer testing and feedback and will not graduate to a new stable version. For applications requiring stability, developers should continue to use the stable gemini-2.5-flash and gemini-2.5-flash-lite models.