AI News Feed
Google Unveils Gemini 2.5 Browser-Control AI
09 Oct 2025- Google’s Gemini 2.5 Computer Use—browser‑control AI—outperforms rivals in Browserbase tests, faster and more accurate, automates UI actions with safety confirmations; available to developers via API preview.
09 Oct 2025
Google has released Gemini 2.5 Computer Use, a browser-control variant of Gemini that third-party Browserbase testing found often “50% faster,” more accurate, and cheaper than competitors (per Poke.com and Browserbase benchmarks). These computer‑use models take screenshots, decide UI actions (clicks/fills), execute them, then re-check the page to iterate until the task is done — effectively letting an AI “use” the browser to complete workflows.
Browserbase ran 200+ experiments totaling ~4,000 browser hours. On benchmarks like Mind2Web, Google hit 69% success versus Claude Sonnet 4.5 at 53% and OpenAI at 46%; on WebVoyager (multi‑step tasks) Google also outperformed the others and maintained lower latency. Google says it trained specifically for pixel precision and optimized for parallel actions (doing multiple steps concurrently) to reduce misclicks and speed up execution.
Google has built safety guardrails that require human confirmation for risky actions (purchases, bypassing CAPTCHAs). Internally it’s already used to fix UI tests (recovering 60%+ of failures) and in Firebase testing, Project Mariner, and AI Mode in Search. Developers can try the Browserbase demo environment and build agents via the Gemini API (model gemini-2.5-computer-use-preview-10-2025) with an execution loop using tools like Playwright.
Source
Google has released Gemini 2.5 Computer Use, a browser-control variant of Gemini that third-party Browserbase testing found often “50% faster,” more accurate, and cheaper than competitors (per Poke.com and Browserbase benchmarks). These computer‑use models take screenshots, decide UI actions (clicks/fills), execute them, then re-check the page to iterate until the task is done — effectively letting an AI “use” the browser to complete workflows.
Browserbase ran 200+ experiments totaling ~4,000 browser hours. On benchmarks like Mind2Web, Google hit 69% success versus Claude Sonnet 4.5 at 53% and OpenAI at 46%; on WebVoyager (multi‑step tasks) Google also outperformed the others and maintained lower latency. Google says it trained specifically for pixel precision and optimized for parallel actions (doing multiple steps concurrently) to reduce misclicks and speed up execution.
Google has built safety guardrails that require human confirmation for risky actions (purchases, bypassing CAPTCHAs). Internally it’s already used to fix UI tests (recovering 60%+ of failures) and in Firebase testing, Project Mariner, and AI Mode in Search. Developers can try the Browserbase demo environment and build agents via the Gemini API (model gemini-2.5-computer-use-preview-10-2025) with an execution loop using tools like Playwright.
Source
The method
The prompts
Helping entrepreneurs and business owners use AI tools for growth and productivity.
Wawa Llama Ltd GB447340593