From Prompt to Page: Gemini 2.5 Takes the Wheel


Google has rolled out Gemini 2.5 Computer Use, a specialized model that can see what’s on a screen and act inside a web browser—clicking, typing, scrolling, dragging, and submitting forms to complete multi-step tasks end-to-end. Rather than relying on each site’s APIs, the model navigates pages visually, much like a human tester or assistant would. Google says the feature is available in preview for developers, with access through Google AI Studio and Vertex AI. 


Under the hood, Computer Use runs as a tight perception-action loop. Your app sends a fresh screenshot (and URL) to the model; the model returns the next UI action to take—such as “click this button,” “type that text,” or “drag to this region.” The client executes the action, captures a new screenshot, and repeats until the task is finished or a guardrail triggers. This approach lets agents handle brittle, real-world UI quirks (modals, cookie prompts, CAPTCHAs you handle externally) without bespoke integrations for every site. 


Google emphasizes scope and safety. The initial release is optimized for browsers (not full desktop OS control), with promising early results on mobile UI tasks as well. A published model card notes inherited limitations from Gemini 2.5 Pro (e.g., occasional hallucinations or shaky causal reasoning) and outlines red-team testing aligned with Google’s AI Principles. In short: useful today, but still bounded and monitored. 


For builders, Computer Use lands at a sweet spot: it’s as straightforward to wire up as function calling, but it opens far richer automations—think customer-support workflows that log into vendor portals, QA bots that execute step-by-step UI tests, or internal tools that reconcile orders across sites that don’t expose APIs. Google’s docs spell out the action schema and developer flow, and the Gemini API changelog pegs the preview launch to October 7, 2025 (India time: Oct 8). 


Early coverage highlights the practical angle: because the model interacts through a browser, it can operate on services where API access is limited or unavailable—useful for automating tedious form-fill tasks or validating flows in staging. Reporters also note Google’s list of predefined actions (open, type, click, drag, etc.), which maps cleanly to common web interactions and makes behavior more predictable for developers. 


If you’ve been waiting for credible, controllable web agents, Gemini 2.5 Computer Use is Google’s answer: a safer, browser-first path to automation that mirrors how people actually get things done online—one click at a time.

Comments

Post a Comment

Popular posts from this blog

OpenAI announces AMD partnership: a 6-gigawatt bet on AI compute

Image
  OpenAI has inked a multi-year, multi-generation agreement with AMD to deploy 6 gigawatts of Instinct GPUs to power OpenAI’s next-gen AI infrastructure. The first tranche— 1 GW using AMD’s Instinct MI450 —is slated to come online in the second half of 2026 , with deployments expanding across future AMD GPU generations.   Beyond raw capacity, the deal includes a notable equity kicker : AMD has issued OpenAI a warrant for up to 160 million AMD shares , vesting as deployment milestones are met (e.g., the initial 1 GW, then scaling toward 6 GW), and subject to additional performance and share-price thresholds. Several outlets report the warrant is structured at a nominal strike price, underscoring the long-term alignment between the companies.   For AMD, leadership says the partnership is expected to drive “tens of billions of dollars” in revenue and be accretive to EPS , a signal of how large-scale AI buildouts are reshaping chipmaker economics. For OpenAI, lo...
Read more

Shields Up: Inside OpenAI’s October Threat Report

Image
  OpenAI’s new October 2025 Threat Report lands with a clear message: malicious actors are experimenting with AI, but mostly to speed up old tricks , not invent new ones. Since beginning public threat reporting in February 2024 , OpenAI says it has disrupted and reported 40+ networks abusing its services, and it continues to ban accounts and share findings with partners when activity crosses policy lines. The post accompanying the report went live on October 7, 2025 (India time: Oct 8 ).   What the investigators are seeing The company’s analysts say threat groups are bolting AI onto existing playbooks —from phishing kits to influence ops—to move faster or write cleaner lures, rather than unlocking novel offensive capabilities. In practice, that often looks like using ChatGPT to plan or draft content , then switching to other models and tools for execution. Some actors even try to mask AI fingerprints by instructing models to avoid tell-tale punctuation patterns. Despi...
Read more

OpenAI brings “Apps in ChatGPT” — and a preview Apps SDK to build your own

Image
  OpenAI has flipped a big switch in ChatGPT: you can now chat with apps like Spotify, Canva, Coursera, Booking.com, Expedia, Figma and Zillow directly inside the conversation. Alongside this, OpenAI is shipping an Apps SDK (preview) so developers can build their own in-chat apps that feel native to ChatGPT.   Why this matters Until now, ChatGPT mostly answered questions or called simple actions. Apps take a step further: services can present rich interfaces and multi-step flows inline with your messages, so you can create a playlist, draft a presentation, enroll in a course, or compare rentals without ever leaving the chat. Early access is rolling out to logged-in users outside the EU across Free, Go, Plus and Pro plans, with EU availability “soon.”   How it works Under the hood, Apps are built on the Model Context Protocol (MCP) —OpenAI’s open standard for connecting models to external tools and data. The Apps SDK extends MCP so developers define bo...
Read more