GPT-6 Astra: The next generation in intelligence for work

What changed

Source summary

OpenAI launched a dedicated enterprise release of GPT-6 Astra, now generally available in ChatGPT Work, Codex, and the API, with computer-use capabilities for apps without APIs, new admin controls to approve which websites/apps it can act on, and new plugins for Oracle Analytics, Power BI, Navan and Avalara. OpenAI says Astra completed a complex financial-modelling benchmark about four times faster than the winning human competitor, and pricing starts at $10 per million input tokens and $50 per million output tokens.

Why it matters to your work

Practical interpretation

If your team uses ChatGPT Work or Codex, this update lets the AI actually operate other apps and websites, including ones without a ready-made integration, to complete tasks like financial reporting or dashboard work in tools such as Power BI or Oracle Analytics instead of just drafting text. New admin approval controls mean IT or compliance can restrict exactly which sites and apps it is allowed to touch, which matters if you are giving it access to systems with client or financial data. It is priced per API token ($10 in / $50 out per million), so it is mainly relevant for orgs already paying for ChatGPT Work or Enterprise or building on the API, not casual ChatGPT users.

When this could help

Worth evaluating for the hardest work tasks — financial modelling, multi-step computer use, agentic coding, cybersecurity analysis — where Sol or earlier models fall short on accuracy or safety.

GPT-6 Astra is OpenAI's most capable model for work. It won the Financial Modeling World Cup 4x faster than the winning human, scored 57.9% on Terminal-Bench 4.0 (vs 37.3% for Sol), and achieved 89% fewer unintended outcomes than Sol on computer use safety. Available via API at $10/M input and $50/M output tokens, with a Fast mode at 2x speed for 2x price.

How to get started

1. In ChatGPT Work or Codex, select GPT-6 Astra as your model. In the API, use model ID gpt-6-astra. For faster responses at higher cost, use the Fast mode endpoint.

   Provider guidance ↗

Build it into your routine

SCALA suggested workflow

1. Model selection for complex work

   Reserve Astra for the hardest tasks — financial modelling, multi-step computer use, agentic coding — where accuracy and safety matter most. Use Sol for routine work to manage cost.

What to check

• At $10/M input and $50/M output, Astra is 5x the price of Sol ($2/$10). Fast mode doubles that again. Source ↗
• OpenAI identifies Astra as reaching the Critical cybersecurity threshold — meaning it can assist with cybersecurity tasks at a level requiring careful access controls. Source ↗