Key Takeaways

  • OpenAI priced GPT-6 Astra at $10 per million input tokens and $50 per million output tokens, targeting high-value enterprise execution over cheap conversational chat.
  • On Automation Bench, Astra doubled the score set by the previous 5.6 Sol model, paired with sharp jumps on Frontier Math Tier 4.
  • The model shifts focus from text generation to direct computer use, executing multi-step tasks inside software like Blender, QA suites, and desktop application builds.
  • OpenAI is staging the release to enterprise customers on Daybreak access first before expanding to Plus and Pro users.

Moving Past Chat to Direct Software Control

Claire Vo evaluated the jump from earlier models like 5.6 Sol and Fable to GPT-6 Astra, and the main shift was not slightly better writing. As Vo put it, “TLDR state of the art on basically everything. We got the science map, coding, knowledge work, not being annoying, being fast, computer use, all the things.”

The central development is computer use. OpenAI built Astra to operate native interfaces, drive external software, and run end-to-end technical tasks. In early testing, that included building CRM logic automatically, generating 3D assets inside Blender, running autonomous QA test suites, handling hardware reverse-engineering, and producing functional desktop applications from a single prompt.

Vo emphasized that software execution is the core design priority of this model: “The big thing that they're calling out is computer use and the use of software.” Instead of handing code snippets back to an engineer to copy, test, and assemble inside an IDE, the system interacts directly with the software environment.

The Economics of Automation Bench and Token Costs

The benchmark data mirrors that practical shift. On Automation Bench, Astra scored twice as high as 5.6 Sol. As Vo noted, “So, automation bench, again, double Soul's reported score. So, you know, a lot higher on automating tasks.” That progress, alongside gains on Frontier Math Tier 4, points to a model tuned for heavy technical labor.

That execution capability comes with higher pricing: $10 per million input tokens and $50 per million output tokens. That rate is steep if you use it for basic copy generation, but OpenAI is positioning the model around labor hours saved.

If an automated run costs a few dollars in output tokens to complete a QA regression pass or configure CRM business rules that used to eat four hours of engineering time, the token expense is minor compared to payroll. The cost only hurts teams that deploy it for tasks simpler models can already handle.

Staged Rollout and Enterprise Access

OpenAI is rolling Astra out in stages rather than opening the gates to everyone at once. “First, it's going to go to enterprise customers with daybreak access,” Vo explained, before the release expands to Plus and Pro subscribers.

This release path shows who OpenAI expects will find immediate value in Astra. Enterprise organizations with structured software pipelines and large developer budgets have the repetitive workflows that justify fifty-dollar output tokens. For startup founders, this creates a clear timeline: identify the most expensive software bottlenecks in your current pipeline now, so you are ready to automate them once general access arrives.

What to Do With This

Audit your five most time-consuming software workflows this week, whether that is QA test runs, CRM routing configurations, or repetitive UI builds. Calculate the exact developer hours spent on each one. When Astra reaches your account tier, test running one entire workflow end-to-end through automated computer use instead of using an LLM as a basic autocomplete tool.