OpenAI's GPT-6 Astra Claims It Can Replace Your PC — Here's What Actually Changes

The claim that changes everything

OpenAI did not quietly update its model lineup on September 3. It released GPT-6 Astra with a single, audacious headline: the system can do anything a human can do on a personal computer. The claim arrived dressed in benchmark data, but the real signal is in the pricing and the rollout strategy — both designed to test whether enterprises will fund the world’s first general-purpose computer agent.

The numbers are striking. On Terminal-Bench 4.0, which measures how well a model navigates actual computer terminals, Astra scored 57.9%. On OSWorld 2.0, a desktop-operation benchmark, it reached 72.6%. But the headline figure is the throughput improvement: average task completion dropped from roughly 75 minutes to 40 minutes. That is the difference between a tool you consult and a worker you assign.

How Astra compares where it counts

Astra posted state-of-the-art results across a wide set of benchmarks. FrontierMath Tier 4 landed at 97.6%. BenchCAD, which tests 3D modeling capability, reached 95.9%. ScreenSpot-Pro — the measure of whether an AI can correctly identify and click UI elements on a high-resolution screen — hit 92.7%. BrowseComp came in at 91.5%, and SRE-Bench, which evaluates the ability to reverse-engineer binary software logic, sat at 88.0%.

The most talked-about metric may be the one that sounds least impressive on paper. AutomationBench, OpenAI’s measure of business-process automation, returned 41.4%. That number looks modest until you factor in the cost side. OpenAI is claiming Astra undercuts competing models on API spend while completing automation tasks faster — a dual advantage that matters far more to a procurement team than a raw percentage.

The ARC-AGI-3 score of 99.9% deserves a separate note. This benchmark tests generalization on novel interactive puzzles, and near-perfect performance suggests the model is learning transferable reasoning strategies rather than pattern-matching to training data. That is the definition of agentic potential.

The safety paradox at the center of Astra

OpenAI built two stories into this launch, and they pull in opposite directions. One is about alignment. On the ExploitGym honeypot — a test that measures whether a model will misuse unauthorized system access to accomplish a task — GPT-5.6 Sol scored a 48.2% exploitation rate. Astra scored zero. OpenAI calls this evidence of superior alignment. The data supports that claim, though it comes from a safeguard-stripped version of the model, which means the real-world figure could shift once production safeguards are reinstalled.

The second story is about capability in exactly the domain where alignment matters most. OpenAI stated that Astra crossed the “Critical” threshold in its own Preparedness Framework for cybersecurity. That means the model can both identify zero-day vulnerabilities and construct exploit code. The company insists operational safeguards prevent actual offensive use, but it is simultaneously rolling out an unrestricted variant through a program called OpenAI Daybreak over the coming weeks for code review and vulnerability analysis. That trajectory — from contained research to wider internal testing — should give regulators pause.

Who pays, and who gets left out

Here is where the announcement diverges from the usual AI-launch playbook. Astra is included in existing ChatGPT Plus, Pro, Business, and Enterprise subscriptions with no separate quota. You can allocate 100% of your usage toward it. That is a deliberate choice. OpenAI is betting that once users experience agentic workflows, they will consume far more than their current allotment — and the upgrade path is already visible.

For API customers, the pricing tells a different story. The Standard tier costs $10 per million input tokens and $50 per million output tokens. Cache reads are $1 per million; cache writes are $12.50. A Fast mode, claimed to be 2.5 times quicker, doubles the price. Zero data retention is available for API subscribers. Amazon Bedrock will also carry the model.

The asymmetry is notable. Consumer subscribers get unlimited access within their plan. Enterprise buyers face per-token costs that scale linearly with usage. If Astra truly replaces desktop workflows, the API bill for a mid-size engineering team could exceed six figures annually within months. OpenAI appears to expect that outcome and is pricing accordingly.

What this means for Anthropic and the rest

Anthropic has staked much of its positioning on being the responsible alternative to OpenAI. Astra’s zero-exploit rate undercuts that argument directly. Claude Fable 5.1 is listed alongside Astra in OpenAI’s comparison charts, but the benchmarks do not show Claude matching Astra’s computer-use scores. If enterprises begin evaluating agents on terminal navigation and desktop interaction rather than text generation, Anthropic faces a capability gap that branding alone cannot close.

Google and Microsoft are in a similar bind. Their computer-use work remains fragmented across products, and neither has announced a unified model that combines reasoning, interface manipulation, and security analysis at this level. Microsoft’s Copilot ecosystem is the closest contender, but it lacks the raw benchmark dominance Astra is displaying.

The rollout reveals OpenAI’s confidence

Astra began limited organizational access on September 3 and is scheduled for broader ChatGPT availability within days. That speed is unusual for a flagship model and signals OpenAI’s belief that production failures will be manageable. The company is also offering a restricted-but-powerful variant through Daybreak, which functions as both a safety valve and a early-warning system for the kinds of autonomous behavior that occur when agents touch real systems.

The subtext is clear: OpenAI expects Astra to change how people use computers, and it wants feedback from enterprises before the general release creates irreversible dependency patterns. The pricing structure rewards early adopters while ensuring that high-volume users subsidize the platform’s continued development.

What happens next

Watch three things. First, the Daybreak expansion schedule — if unrestricted Astra variants reach a broad developer audience, the cybersecurity implications become concrete, not theoretical. Second, enterprise API bills — the first public reports of real-world token consumption will reveal whether Astra’s automation claims translate into cost savings or cost explosions. Third, regulatory reaction — an AI that can both find and exploit zero-days while claiming perfect alignment will attract attention from the FAA-level oversight bodies already discussing autonomous software.

Astra is not the end state of agentic AI. But it is the first model that makes the transition from assistance to automation economically plausible for a company willing to pay the token bill.