OPENAI / MODEL PROFILE
GPTComplex work, carried through.
GPT-6 Astra
A candidate for projects that combine research, analysis, coding, and work across software tools.
Research checked September 5, 2026
THE ASSESSMENT
Documented capabilities and limitations, with practical interpretation.
Tokens are small units used to represent inputs and outputs. API rates are usage charges for software, separate from chat subscriptions. Prices are in US dollars, as checked for this edition.
Read our research method →Where it stands out
Work across several steps
OpenAI reports improvements in computer use, browsing, software engineering, and professional tasks. The practical attraction is continuity: assembling information, working with it, and producing a deliverable in one supervised workflow.
OpenAI · Astra launch and evaluations ↗Stronger factuality, with evidence
Its system card reports fewer factual errors than GPT-5.6 Sol on selected user-flagged conversations. Those deliberately difficult tests are evidence of improvement, not a promise that everyday answers will be correct.
OpenAI · Astra system card ↗Steer work while it runs
The API supports additional instructions during a running task and asynchronous tool calls. These features can help a custom application respond to corrections while other work continues; the application must implement them.
OpenAI · Model behavior and prompting ↗Where to be careful
It can overwork a simple request
OpenAI documents a tendency toward detailed formatting, extra testing, and clarification questions. A short email or routine summary may need a tighter brief, or a less demanding model.
OpenAI · Model behavior and prompting ↗Premium pricing and interruptions
Standard API pricing is $10 per million input tokens and $50 per million output tokens. Safety checks can pause legitimate work. Measure the cost of a completed, reviewed task, including retries, rather than assuming the newest model is the economical choice.
OpenAI · Astra launch and evaluations ↗Better reasoning still needs checking
The system card retains factual-error and alignment limitations. A confident explanation is insufficient evidence for a citation, a calculation, or an action taken in another application.
OpenAI · Astra system card ↗OUR SUGGESTED TRIAL
Try it on your kind of work.
Give it a fictional vendor proposal and a small cost spreadsheet. Ask for a one-page decision memo, a calculation table, and an explicit list of missing facts. Require a source reference for each factual claim.
What to look for
Compare the final numbers with your own calculation. Count invented assumptions and the time you spend shortening or correcting the memo.
An editorial test idea, not a reported benchmark result. Use public, fictional, or properly approved material.What else to compare
Keep GPT-5.6 in the comparison for routine work. OpenAI’s own guidance treats task cost and model behavior as separate considerations from maximum capability.
OpenAI · Model behavior and prompting ↗No model in this guide replaces source checking, professional judgment, or a review of the application’s data terms.
Sources & further reading
Primary documentation checked for this edition. Provider claims are not an independent head-to-head test.