GPT-6 Astra: 72.6% Computer Use, $10/$50 Pricing, Critical Cyber Tier

OpenAI shipped GPT-6 Astra on September 5, 2026 — less than a week after Anthropic released Claude Fable 5.1 — positioning it as its most capable and aligned model to date. For developers and ML practitioners, the release is notable less for raw benchmark numbers and more for what those numbers represent: a model architected around autonomous task completion rather than assisted generation. Astra can operate a desktop end-to-end, produce deliverables that match an organisation's existing templates, retain searchable state across long coding sessions, and — most consequentially — independently develop working exploits for previously unknown vulnerabilities, a capability OpenAI's own Preparedness Framework classifies at its highest risk tier.

Computer Use and Agentic Judgment

The most concrete signal in Astra is its OSWorld 2.0 score of 72.6%, which measures real desktop operation — form completion, CRM updates, frontend QA, software troubleshooting — without step-by-step guidance. That places Astra ahead of Claude Opus 5's 70.2% and well ahead of GPT-5.6 Sol's 65.7%. OpenAI's own demo illustrates the judgment layer: GPT-5.6 Sol autonomously built a personal career website in approximately 13 minutes, while Astra paused after 20 seconds to ask which career the user was transitioning into before proceeding. That pause-vs-act distinction reflects deliberate training — the model fills routine gaps independently and interrupts only when the answer would materially change the outcome. For teams building production agent pipelines, that calibration matters more than throughput in most agentic workflows.

Astra's memory architecture in Codex extends this further. Rather than compressing long debugging sessions into a rolling summary — which typically discards context like the rationale behind a failed fix — Astra maintains searchable notes across context windows. The feature is opt-in at launch via Codex's config file, with OpenAI stating it will become the default in the coming weeks.

Benchmarks

All figures below are self-reported by OpenAI in its launch materials.

Astra's leads on FrontierMath Tier 4 (97.6% vs Claude Fable 5.1's 87.8%) and Terminal-Bench 4.0 (57.7% vs Fable's 55.8%) are real but asymmetric in size. The Humanity's Last Exam gap runs the other direction: Astra's 57.2% trails both Claude Fable 5.1 at 65.0% and Claude Opus 5 at 63.6%. One headline figure warrants scrutiny: the marketed 99.9% on ARC-AGI-3 depends on an expensive, stateful evaluation harness. Independent testing by the ARC Prize Foundation put standard stateless API performance between 17% and 63% depending on the reasoning tier — teams running normal API integrations should calibrate to the lower bound.

Cybersecurity Capability and Access Gating

Astra's 100% score on ExploitBench and 88% first-attempt success rate on SRE-Bench reverse-engineering tasks are the numbers that drew the most scrutiny at launch. Those scores placed the model at OpenAI's "Critical" cybersecurity tier under its Preparedness Framework — the first time any OpenAI model has crossed that threshold — meaning it can independently identify and build working exploits for previously unknown vulnerabilities. The governance question this raises is not hypothetical: OpenAI is gating exploit-creation features behind its Daybreak program at launch, while permitting defensive applications like secure code review and patch validation without restriction.

Pricing and Rollout

API pricing for gpt-6-astra is $10 per million input tokens and $50 per million output tokens. A fast mode running at roughly 2.5× standard processing speed is available at 2× the price. Those rates sit substantially above GPT-5.6 Terra's $2/$12 and Claude Opus 5's $5/$25. The model is available through the OpenAI API, Microsoft Azure, and Amazon Bedrock. On the consumer side, it is rolling out to ChatGPT Plus, Pro, Business, and Enterprise users in stages; Enterprise admins must manually enable it per workspace since it defaults to off. Pro, Business, and Enterprise subscribers also receive access to a separate GPT-6 Astra Pro tier. Zero data retention is available for eligible API customers, and OpenAI says it is testing private safety processing to strengthen monitoring while preserving customer confidentiality.

The $10/$50 per million token pricing is not positioned for bulk text generation. It targets workflows where a task completes reliably with minimal supervision — the same space where infrastructure governance determines whether autonomous agents can actually be trusted at production scale. Astra's benchmark profile, particularly its computer use scores and persistent memory in Codex, suggests OpenAI is betting that a subset of enterprise workloads will generate enough value per completion to justify the premium. Whether that holds will become visible once the broader rollout reaches Enterprise customers and independent red-teaming begins.