COMPUTER USE
Can Astra use a computer like a person?
OpenAI documents computer use; the hard part is measuring it in your environment.
GPT-6 Astra’s official model guidance lists computer use as a supported capability. This page keeps that product fact separate from the older reported human-parity phrase and the independent task battery still to run.
CURRENT EVIDENCE
Documented capability · independent score pending.
OpenAI’s current model guidance lists computer use among Astra’s capabilities. Sources News separately reported a roughly human-parity description; that phrase remains context, not a defined benchmark result.
WHAT THE RECORD SAYS
The headline is useful; the conditions are missing.
“Computer use” can mean anything from visual navigation to long-running work across multiple applications. The public record does not yet tell a reader which version was demonstrated or how it handled mistakes.
Reported capability
ReportedComputer use is listed as a supported GPT-6 Astra capability.
The official model record establishes support; it does not establish task reliability.
Public access
ConfirmedAstra is released; access routes are expanding.
ChatGPT, Work, Codex, API, Copilot, Foundry, and catalog routes have separate eligibility terms.
Permissions
UnknownBrowser, desktop, filesystem, network, and approval boundaries are unknown.
Never infer permissions from a headline about capability.
Reliability
UnknownNo independent error, recovery, latency, or cost distribution is available here yet.
A successful demonstration would not establish dependable unattended use.
INDEPENDENT TEST PLAN
What we will test when access exists.
The task battery is designed for inspectable behavior, not a single impressive demo. Every run needs a reversible environment and a human approval boundary.
Navigate
Complete bounded, multi-step tasks in a sandbox with a fixed application set.
Recover
Measure whether the agent notices a wrong state, explains it, and returns to a known checkpoint.
Respect
Block credential access, irreversible actions, and destinations outside the explicit allowlist.
Disclose
Publish artifacts, failures, time, cost, retries, and the tasks the result does not cover.
DOES NOT PROVE
A strong claim is not a deployment policy.
Until the conditions are public, the responsible conclusion remains narrow.
- It does not prove every ChatGPT plan, API account, region, or desktop environment has access.
- It does not establish safe unattended control of a computer or access to a reader’s data.
- It does not provide a general comparison with GPT-5.6 Sol, Claude, Gemini, or a human operator.
- It does not replace the published system card or a reproducible public task battery.
SOURCE LEDGER
Reported claim, official context.
OpenAI’s model documentation supplies the current computer-use capability record. The reported human-parity phrase remains separately labeled, and the independent task protocol is still unrun.
FAQ
Astra computer-use questions
Can I use Astra to control a computer today?
OpenAI’s current model guidance lists computer use as a supported GPT-6 Astra capability, but access is rolling out in stages and no independent public task battery is available on this site yet.
What does “human parity” mean here?
The older “human parity” phrase is not a defined score. It does not tell us which applications, tasks, error rates, permissions, or recovery behavior were tested.
What would make the claim testable?
A frozen task set, documented environment and permission policy, raw artifacts, failures, cost, latency, and independent review would turn a capability claim into evidence.
ASTRA UPDATES
Keep the useful Astra updates coming
We’ll send occasional source-backed updates as access, pricing, and real-world tests change. This is an independent list, not an OpenAI waitlist.