Google DeepMind has pushed its Gemini 2.5 models – 2.5 Pro and the lighter 2.5 Flash—into general availability across Vertex AI and Workspace after a month-long preview. The upgrade delivers native audio output, tighter security guards and Deep Think, an optional mode that chains hundreds of reasoning steps to crack complex code and math in a single prompt. Benchmarks released at I/O show 2.5 Pro overtaking GPT-4o on MMLU and GSM-8K, while using 30 % fewer tokens thanks to a revamped retrieval pipeline.

Developers can now hot-swap between Flash – optimised for latency-sensitive workloads under 300 tokens—and Pro for heavyweight tasks without touching their front-end code. Google also widened rate limits, slashed Pro pricing by 25 % and opened sign-ups for Gemini Live, a multimodal assistant slated to replace Google Assistant on Android later this year. Privacy advocates praise on-prem inference options, while enterprise CIOs eye Pro’s new Project Mariner abilities that let agents navigate desktop UIs to automate invoices or design mock-ups. Analysts at IDC predict Gemini’s tiered release will pressure rivals to roll out similar “reasoning boosts” before year-end.