OpenAI GPT-6 Astra: Enterprise Computer Use in the AGI Era
On September 3, 2026, OpenAI released GPT-6 Astra, its first GPT-6 class model, two months after GPT-5.6 Sol and days after Anthropic released Claude Fable 5.1 and Mythos 5.1. OpenAI president Greg Brockman closed the press briefing with: “Welcome to the AGI era.” The framing is deliberate, but the enterprise value proposition is concrete: a model built to operate computers on behalf of employees, complete multistep professional workflows without manual intervention, and pass OpenAI’s most stringent internal safety review to date.
What GPT-6 Astra actually does
OpenAI is positioning Astra around computer use as the primary differentiator. The model can navigate browsers, fill forms, update customer records, build and test websites, create documents and spreadsheets from company templates, and execute multiple tasks across applications simultaneously. A launch video showed an employee asking Astra (by voice) to turn a sketch into a 3D model, then build a game from it, while simultaneously placing a food order. OpenAI calls it “the world’s best computer use model.”
That claim is backed by benchmark results. On OSWorld 2.0, which tests autonomous computer operation on real software, Astra scores 72.6% versus 65.7% for GPT-5.6 Sol. Average task completion time drops from 75 minutes to about 40 minutes, a 47% reduction. On ScreenSpot-Pro (finding and clicking the correct UI element in dense interfaces), Astra scores 92.7% against 76.9% for Sol.
AutomationBench, which measures end-to-end professional workflow completion, shows the widest gap in the announcement: Astra at 41.4% versus Fable 5.1 at 31.4% and Sol at 18.1%. BenchCAD (reconstructing CAD programs from rendered views) goes to Astra at 95.9%, versus 84.3% for Fable 5.1 and 83.3% for Sol. OpenAI notes the Fable 5.1 comparison used modified evaluation settings; independent replication is still pending.
For coding, the Codex component adds cross-window memory: Astra can retain notes and search earlier context rather than relying on compressed summaries. This matters for large codebases where critical detail gets lost as sessions grow. On Terminal-Bench 4.0, Astra scores 57.9% versus Fable 5.1 at 55.8% and Sol at 37.3%. The Artificial Analysis Coding Agent Index is effectively a three-way tie (Fable 5 at 68.1, Fable 5.1 at 67.2, Astra at 67.0).
Benchmark summary
| Category | GPT-6 Astra | Claude Fable 5.1 | GPT-5.6 Sol |
|---|---|---|---|
| AutomationBench | 41.4% | 31.4% | 18.1% |
| OSWorld 2.0 | 72.6% | n/a | 65.7% |
| Terminal-Bench Science 0.1 | 64.6% | 52.6% | 22.4% |
| Terminal-Bench 4.0 (coding) | 57.9% | 55.8% | 37.3% |
| FrontierMath Tier 4 | 97.6% | 87.8% | 83.0% |
| GPQA Diamond | 96.0% | 93.7% | 94.6% |
| ARC-AGI-3 | 99.9% | n/r | 7.8% |
| ExploitBench | 100% | 70.0% | 78.5% |
| Humanity’s Last Exam (w/ tools) | 57.2% | 65.0% | n/r |
| AA Intelligence Index v4.1.1 | 61.2 | 65.7 | 60.9 |
Where OpenAI’s numbers lead, they do so clearly. Where they do not, Astra often omits the benchmark from launch prose. Humanity’s Last Exam is the clearest example: Astra’s 57.2% trails Fable 5.1 (65.0%), Fable 5 (63.8%), and Opus 5 (63.6%). The Artificial Analysis Intelligence Index, run by a third party, puts Fable 5.1 four points ahead. Enterprise teams should run their specific workloads rather than rely on any single table.
Critical cybersecurity threshold: what it means
GPT-6 Astra is the first model to reach the “Critical” level under OpenAI’s Preparedness Framework for cybersecurity. On ExploitBench, Astra scores a perfect 100% (turning known vulnerabilities into working exploits) versus 78.5% for Sol. On a contamination-controlled internal test using 20 high-severity V8 vulnerabilities disclosed June through August 2026, Astra achieves 39% arbitrary code execution versus Sol at 5.5%. During testing it discovered and used two previously unknown zero-day vulnerabilities, which OpenAI is disclosing to maintainers.
The Critical rating triggered enhanced rollout controls. The version available to general enterprise accounts refuses proof-of-concept exploit generation. Full capabilities are gated behind the Daybreak program for vetted cybersecurity defenders, where Astra can assist with vulnerability validation, malware analysis, and detection engineering.
For security teams this creates a clear evaluation question: does your organization qualify for Daybreak access, and if so, does Astra’s SRE-Bench result (88.0% solve rate versus 55.9% for Sol) justify the access process? For CISOs not in Daybreak, the practical change is updating AI acceptable-use policies to reflect that the model can identify attack paths even without the unrestricted capability tier.
Safety architecture
Alongside raw capability, Astra introduces structural changes to alignment. On the internal computer-use safety benchmark (lower is better), Astra scores 2.4% versus 22.0% for Sol. With AutoReview enabled, the rate drops to 1.8%. On the ExploitGym honeypot benchmark (going beyond authorized scope), Sol registered 48.2% in testing; Astra registered 0.0%. Internal hallucination rates drop from 12.2% to 4.2%.
OpenAI built 24/7 escalation monitoring into the deployment, with a target response time of 30 minutes for potential misalignment signals. One caveat in the launch materials: Astra’s written reasoning was harder to monitor than Sol’s in tests designed to detect monitoring evasion. OpenAI describes improving “monitorability” as an active research priority, not a solved problem.
The model also went through the White House’s voluntary vetting process for frontier AI systems, per OpenAI CEO Sam Altman.
Pricing and availability
| Tier | Input | Output |
|---|---|---|
| GPT-6 Astra (standard) | $10.00 / 1M | $50.00 / 1M |
| GPT-6 Astra Fast mode | $20.00 / 1M | $100.00 / 1M |
| GPT-5.6 Sol (promo) | $4.00 / 1M | $20.00 / 1M |
| Claude Fable 5.1 | $10.00 / 1M | $50.00 / 1M |
Fast mode runs at up to 2.5x standard speed. Astra costs 2.5x more per token than Sol at Sol’s current promotional price. Parity with Fable 5.1 on the token price means the contest shifts to task completion cost: OpenAI claims Astra uses approximately 70% fewer output tokens than Sol on coding tasks and roughly 65% fewer than Fable 5.1 on some professional benchmarks.
Artificial Analysis found that on the Coding Agent Index, Astra equals Fable 5 at less than half the per-task cost, driven by token efficiency. On the Intelligence Index at max effort, Astra is about 75% more expensive per task than Sol, because the price increase outweighs the token savings. The calculus depends entirely on workload type.
Availability is staged: Daybreak enterprise customers first, then paid ChatGPT plans (Plus, Pro, Business, Enterprise) and the API over the coming days. AWS Bedrock (model name: gpt-6-astra) and Microsoft Azure are supported at launch. Enterprise workspace admins must enable the model manually; it is off by default. Zero Data Retention is available for eligible API customers. OpenAI is testing Private Safety Processing as an additional governance option.
There are no Luna/Terra/Sol variants for GPT-6. The lineup is Astra (standard) and Astra Pro (for Pro, Business, and Enterprise plan holders).
What enterprise teams should do now
For most enterprises, the near-term action items are straightforward.
First, update AI acceptable-use policies. The model’s Critical cybersecurity rating is the first of its kind, and your policies should reflect that the base model (even with restricted cyber capabilities) is more capable at vulnerability discovery than prior generations.
Second, identify computer-use workloads. AutomationBench’s 41.4% result is the clearest differentiation from anything previously available. Document-heavy workflows, web research, and form-filling pipelines are the immediate candidates for evaluation. Anthropic’s computer use and Skills API reached general availability in August 2026, so this is a live comparison for enterprise architecture teams.
Third, wait for broader availability before committing to cost estimates. Launch benchmark data does not settle cost-per-task comparisons across production workloads. Run your own evals against your own repositories and workflows before assuming the token efficiency story holds.
Fourth, if your organization runs a security program that would qualify for Daybreak, evaluate that access path. The ExploitBench and SRE-Bench results represent a step change in what AI can do for defensive security, not just a marginal improvement.
OpenAI is positioning Astra as IPO-era proof that enterprise buyers should choose OpenAI for their most important workflows. Whether the benchmarks hold in production depends on the same thing they always have: the specific task, the specific codebase, and the specific team running the evaluation.
Sources: The Verge, VentureBeat, OpenAI benchmark page, Artificial Analysis, Bloomberg, Vellum AI, NBC News