GPT-6 Astra: $10/M Tokens, 57.9% Terminal-Bench, Critical Cyber Flag

OpenAI last week shipped GPT-6 Astra, now available across ChatGPT Work, Codex, and the API. The release's structural addition is native computer use: the model can operate desktop applications without requiring those applications to expose an API. Instead of building custom connectors, teams can point Astra at existing software and have it work through interfaces the way a human operator would — compressing or eliminating the usual pre-deployment runway of data preparation, workflow redesign, and bespoke tooling.

OpenAI's own engineering team demonstrated this internally before launch, using Astra to identify and fix a memory-allocation bottleneck in Codex test environments. Switching allocators based on Astra's analysis produced 25× lower turn latency at the cost of roughly 30% higher peak memory use — a real engineering trade-off, not a marketing abstraction.

Computer Use and Task Efficiency

On Terminal-Bench 4.0 — which tests agents on terminal-based software engineering, system configuration, and data analysis — Astra scores 57.9%, against 37.3% for GPT-5.6 Sol 2 and 55.8% for Claude Fable 5.1. The cost delta is material: Astra runs at approximately 9% lower estimated API cost per task than Claude Fable 5.1, and approximately 63% lower than GPT-5.6 Sol 2 on that benchmark.

Datacurve reports Astra set a new record on DeepSWE v1.1 at 74%, achieved with fewer steps and greater token efficiency than prior frontier models. CodeRabbit found Astra caught roughly 20% more bugs than their baseline, and more than doubled the catch rate on pull requests requiring extensive cross-file reasoning. Basis reports a 20% improvement in pass rate for end-to-end agentic workflows spanning five or more hours, alongside a reduction in the total number of inference calls needed — a direct indicator that the model completes tasks with less scaffolding overhead, which matters for infrastructure governance around safe agent deployment.

API Pricing and Cost Efficiency

Pricing is set at $10 per million input tokens and $50 per million output tokens. OpenAI claims Astra occupies a majority of the cost-efficiency frontier on professional work and coding evaluations, including both Terminal-Bench 4.0 and the Artificial Analysis Intelligence Index. The model was trained to complete tasks in fewer tokens with fewer retries, reducing per-task cost in production workflows.

Databricks confirmed Astra offers significantly better cost per task than GPT-5.6 Sol on OfficeQA Pro and Pro V2 benchmarks, with clear gains in data reasoning and document understanding. Hebbia added a precision dimension: Astra followed analyst briefs 17% more faithfully than the next-best model and sourced claims to the correct document 19% more often, reducing the verification burden on human reviewers.

Safety Architecture and the Preparedness Threshold

Astra is the first model OpenAI has flagged as reaching the Critical cybersecurity capability threshold under its Preparedness Framework — a designation that triggered hardened controls rather than a release hold. On OpenAI's internal computer use safety benchmark, which tests against scenarios including exposing confidential information, over-sharing dashboards, and deleting data, Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1. Box's evaluation found Astra was more than 10% less likely to make confidently incorrect assertions compared to their baseline.

New enterprise admin controls extend that safety posture to deployment configuration: administrators can restrict access to approved websites and desktop applications, manage file transfers, and control browsing history. Confirmation policies can gate consequential actions behind human approval, and automated review applies to potentially unsafe or unauthorized tool calls. Enterprise access is off by default at launch, managed under existing rate cards and agreements. Zero Data Retention is available for eligible API customers on supported endpoints, subject to approval.

Astra's native computer use collapses a layer of integration engineering that has historically separated AI capability from enterprise deployment — the same friction point that production AI failures so often trace back to architecture rather than model intelligence. Paired with the Preparedness Framework threshold disclosure, OpenAI is shipping capability gating and safety instrumentation together rather than sequentially. Developers building on the API should expect that pattern — higher capability paired with more granular control surfaces — to continue with future flagship releases.