$ today_in_10_seconds
OpenAI unveiled GPT-6 Astra and is calling it the world's best computer-use model, with tests showing it booking DMV appointments and searching job listings faster than a human. It's priced at $10/$50 per million tokens, identical to Claude Fable 5.1. The frontier is now openly competing on autonomy, not chat.
$ lead_story
GPT-6 Astra makes computer use the headline feature
OpenAI's next-generation model is being sold on one capability: operating a computer like a person does. In internal tests it navigated real-world web bureaucracy (DMV scheduling, job boards) end to end, and OpenAI leadership is framing it as a milestone rather than an increment. That's a shift in what a flagship launch means: coding and computer use are the marquee benchmarks, and chat quality is barely mentioned. For builders, the interesting part is that the DMV task is a messy, multi-step, click-and-wait workflow with no API, which is exactly the class of problem agents have historically failed at.
// why_it_matters: When the frontier lab's launch demo is a DMV form and not a poem, the market for thin-wrapper browser agents just got a lot harder.
$ also_worth_attention
Astra's pricing lands exactly on Claude Fable 5.1: $10 in, $50 out
OpenAI matched Anthropic's per-token pricing to the dollar. For agentic workloads, where a single task can burn hundreds of thousands of tokens across tool calls and screenshots, the real cost differentiator is now tokens-per-task-completed, not the rate card. Price parity means the benchmark that matters is efficiency.
link →The Hugging Face breach model ships with new safeguards
Astra is the same model that leaked in the Hugging Face breach, and OpenAI is shipping it anyway, with added protections. Expect those safeguards to bite hardest on computer-use tasks (payments, credentials, account changes), since that's where an autonomous model can do real damage. Builders should assume more confirmation prompts and refusals in exactly the flows they most want automated.
link →"Faster than the average person" is the new benchmark
OpenAI compared Astra to humans on task completion time, not to other models on a leaderboard. That's the metric enterprises actually buy on, and it signals where evals are heading: wall-clock time to finish a real job, including recovering from broken pages and CAPTCHAs.
link →
$ bigger_picture
The trend behind the headlines
Two things converged today: the top labs now lead with autonomy, and they've stopped competing on price. Identical pricing between OpenAI and Anthropic pushes the fight to reliability, tool-use efficiency, and how much supervision a model needs before you'll let it touch your accounts. The DMV demo is a tell: the target isn't developers writing code, it's every knowledge worker with a browser and a tedious task. Companies building agent products on top of these models should expect the base capability to keep eating their differentiation, and invest in the layers the labs won't: domain data, permissions, audit trails, and recovery when a run goes wrong.
Browser Use
An open-source Python library that gives any LLM a structured browser (DOM extraction plus screenshots) to complete multi-step web tasks. If you want to test how Astra, Fable, or an open model handles the kind of forms-and-clicks workflows in today's demo, this is the fastest harness to spin up, and it's model-agnostic, so you can benchmark tokens-per-task across providers.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime