OpenAI Pushes GPT-6 Astra Into Enterprise Work, Priced From $10 per Million Input Tokens

On September 9, 2026 OpenAI published an enterprise post titled "GPT-6 Astra: The next generation in intelligence for work", saying the model it introduced last week (September 3) is now available in ChatGPT Work, Codex and the API, and describing it as "the world's most intelligent and aligned model". Pricing starts at $10 per million input tokens and $50 per million output tokens.

The pitch: no workflow re-plumbing first#

The post argues that most AI systems make businesses prepare data, redesign workflows and build custom integrations before delivering value, whereas Astra "can write code and work through the same applications people use every day — even when those applications don't have an API". In the first days of rollout, customers were already using it to optimize GPUs, spot discrepancies in financial statements and produce more on-brand decks. Alongside Astra, new enterprise plugins arrive in ChatGPT Desktop from Oracle Analytics, Power BI (a Microsoft Fabric service), Navan and Avalara.

What OpenAI and its customers put numbers on#

The figures below appear in the official post; most come from customer evaluations rather than one OpenAI benchmark table:

• Databricks: Astra claims a new state of the art on its OfficeQA Pro and Pro V2 benchmarks using the Genie harness, with significantly better cost per task than GPT-5.6 Sol.
• Hebbia: Astra's decks "followed the brief 17% more faithfully than the next-best model" while sourcing claims to the right document 19% more often.
• Box: across the evaluation Astra was more than 10% less likely to make confidently incorrect assertions.
• Datacurve: Astra set a new record on DeepSWE v1.1 at 74%.
• OpenAI's own account: it used Astra and Codex to find a memory-allocation bottleneck behind slow Codex sessions, and by switching allocators produced 25x lower turn latency with roughly 30% higher peak memory use.

OpenAI says Astra was trained to finish tasks in fewer tokens with fewer retries, and claims it occupies the majority of the cost-efficiency frontier on professional-work and coding evaluations, including Terminal-Bench 4.0 and the Artificial Analysis Intelligence Index.

Safety and control claims#

OpenAI calls Astra the first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework. On an internal computer-use safety benchmark, Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1. New enterprise admin controls can restrict access to approved websites and desktop applications, manage uploads and downloads and control browsing history, alongside confirmation policies and automated review of risky tool calls. Zero Data Retention is available for eligible API customers on supported endpoints, subject to approval.

On the "AGI era" claims#

On ARC-AGI-3, Astra surpassed our human action-efficiency baseline on 96% of levels, effectively reaching human parity on the benchmark. — Greg Kamradt, ARC Prize Foundation, in the September 3 launch post

Coverage of this release has leaned on "AGI era" framing, but in the full text of both OpenAI posts we checked (September 3 and September 9), the term "AGI" appears only inside the names of the ARC-AGI benchmark series; OpenAI's own wording is "the world's most intelligent and aligned model" and it does not claim AGI has been reached. The September 3 post also states that Astra saturates FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9% and ExploitBench at 100%. Pricing, benchmark details and availability should be read from OpenAI's own pages.