Free Google AI 'Gemini 2.5 Flash' will be updated to high speed and high accuracy
Google announced updates to the Gemini 2.5 Flash and Gemini 2.5 Flash-Lite on September 25, 2025. Both models have improved response accuracy and speed, and output costs are lower.
Continuing to bring you our latest models, with an improved Gemini 2.5 Flash and Flash-Lite release - Google Developers Blog
Gemini 2.5 Flash just got a few new updates:
— Google Gemini App (@GeminiApp) September 25, 2025
You'll see enhanced step-by-step help for homework, better-organized responses, and improvements in image understanding. Here's a breakdown of what's new 🧵 pic.twitter.com/Pzv2mYNwKB
Gemini 2.5 Flash is a lightweight, high-performance AI model that can be used with the free version of Gemini, while Gemini 2.5 Flash-Lite is a high-speed AI model developed with a focus on low latency. The graph below compares the performance of Gemini 2.5 Flash (white) and Gemini 2.5 Flash-Lite (blue) before and after the update, with the horizontal axis representing response accuracy and the vertical axis representing response speed. The graph shows that both models are fast and highly accurate.
The graph below compares the number of output tokens for the same task before the update (dark blue) and after the update (light blue). Gemini 2.5 Flash reduced the number of output tokens by 24%, while Gemini 2.5 Flash-Lite reduced it by 50%. Since API usage fees for the Gemini series are charged per token, both models are now cheaper to use.
In addition, Google has listed the following improvements to Gemini 2.5 Flash: 'It is now possible to clearly and step-by-step teach how to solve homework,' 'it is now possible to output complex content in lists and tables,' and 'image recognition accuracy has improved.'
1/ Clearer explanations for homework
— Google Gemini App (@GeminiApp) September 25, 2025
We've improved Gemini's reasoning capabilities for homework questions, so Gemini is much better at walking you through complex subjects with clear, step-by-step guidance.
Additionally, the ' -latest ' alias, which indicates the latest version, has been introduced to the Gemini series API. This means that developers can specify the latest version of each model simply by adding '-latest' to the end of the model name, such as 'gemini-flash-latest.' This eliminates the need to rewrite the Gemini version number in the source code every time Gemini is updated.
There have also been numerous reports of the Gemini series being interrupted more frequently than other AI models, with one posting on the news-sharing website Hacker News stating, 'Gemini's responses are excellent compared to Claude and GPT-4, but the responses are interrupted too often. I would choose a model that is slightly inferior but returns a complete response over a high-performance model that is interrupted mid-response. Until the interrupted response issue is resolved, Gemini will always seem broken, no matter how good its benchmark results.'
Related Posts:
