Google has released upgraded versions of its Gemini 2.5 Flash and Flash-Lite models, improving instruction following, multimodal accuracy, and efficiency while cutting token usage by up to 50 per cent
Google has updated its Gemini 2.5 Flash and Flash-Lite AI models, promising faster, cheaper and more accurate results across a wide range of applications.
In a blog post, the company said the releases are now available in Google AI Studio and Vertex AI, and are designed to balance performance with efficiency.
The latest Flash-Lite update was built on three core themes: better instruction following, reduced verbosity, and stronger multimodal capabilities.
According to Google, the model is now “significantly better at following complex instructions and system prompts,” while also producing shorter, more concise answers.
It also boasts “more accurate audio transcription, better image understanding, and improved translation quality.”
The standard Gemini 2.5 Flash model has also been upgraded.
Google explained: “We’ve improved how the model uses tools, leading to better performance in more complex, agentic and multi-step applications.”
The company highlighted a five per cent gain on SWE-Bench Verified, a key benchmark for software engineering tasks, compared to the previous release.
Efficiency has also been a focus, and Google claims the new Flash and Flash-Lite versions cut output token usage by up to 50 per cent – reducing both latency and cost.
This is particularly aimed at developers deploying large-scale AI systems.
Early testers have reported significant improvements, with Yichao ‘Peak’ Ji, co-founder and chief scientist at Manus, saying: “The new Gemini 2.5 Flash model offers a remarkable blend of speed and intelligence.
“Our evaluation on internal benchmarks revealed a 15 per cent leap in performance for long-horizon agentic tasks.
“Its outstanding cost-efficiency enables Manus to scale to unprecedented levels—advancing our mission to Extend Human Reach.”
Google unveils new improvements to Gemini 2.5 Flash and Flash-Lite
