GPT-6 Astra: Give It a Mission Worth Finishing
It is 4:40 on Friday. You have three supplier documents, a spreadsheet with conflicting totals and a decision due before Monday. That is the assignment we would bring to GPT-6 Astra.
The interesting question is how much of that work survives contact with a real deadline: finding the discrepancy, checking the source, producing an editable comparison and explaining what still needs a decision.
Documentation checked on 5 September 2026. The mission below is our proposed evaluation, not a reported benchmark or customer result. Cover: conceptual illustration created with imagegen.
Astra, in one minute
OpenAI documents Astra as its most capable model for difficult work across code, research and professional tools. Its model ID is `gpt-6-astra`, with a 1,050,000-token context window and up to 128,000 output tokens. Official model details.
Treat the big context window as room for relevant evidence. A folder of obsolete documents still needs dates, owners and a way to tell which version wins.
Two changes you can actually design around
Work can continue during a slow lookup. Async tool calling lets the model proceed while your application runs a tool; your application still owns execution and returning the result. Async tool documentation.
For our supplier exercise, that suggests a useful division: request missing reference material, then inspect the spreadsheet while it arrives. The comparison must still wait for any evidence it depends on. Faster activity only helps when the dependencies are understood.
You can change the brief while it works. Astra supports mid-turn steering through the Responses API over WebSockets. A correction does not undo actions already taken or cancel tools already running. Steering documentation.
Try this halfway through: “Supplier B has withdrawn. Keep its evidence in the appendix, but remove it from the recommendation.” The interesting result is a consistent revision across the summary, calculations and final choice.
Your first mission: the Friday decision pack
Use fictional or authorized material. Include one outdated quote and a deliberately inconsistent total. Define the correct outcome before you start.
Give the model this brief:
Objective: prepare a supplier comparison for a manager who has ten minutes to decide.
Inputs: the attached quotes, spreadsheet and purchasing criteria. Treat the dated criteria as the authority for the recommendation.
Deliverables: an editable comparison, a short recommendation with source references, and a list of unresolved questions. Show the calculation behind every total. Flag missing evidence explicitly.
Working rules: continue through routine choices and record assumptions. Ask when an answer could change the recommendation. Do not send messages, purchase anything or modify source files. Finish with what was checked and what remains uncertain.
This is a reusable brief for a tool-enabled environment, not a claim that every chat interface can read files or create spreadsheets. Confirm the tools and model access in the account you will use.
The price tag deserves a calculator
Standard API pricing is $10 per million input tokens and $50 per million output tokens. Above 272,000 input tokens, the full request moves to long-context rates: $20 input and $75 output. Official pricing.
An illustrative request with 20,000 uncached input tokens and 5,000 total billable output tokens costs $0.45 at the standard short-context rates: $0.20 + $0.25. This arithmetic excludes tools, cache writes, regional uplifts and tax. Count all billable output, including reasoning, rather than just the visible answer. A workflow may make several requests.
Then count your time. If the recommendation takes longer to repair than to write, cheap tokens have not bought a cheap result. Record preparation, waiting, corrections and final review alongside API usage.
Keep score on the deliverable
Evidence: can you open the source behind each material claim?
Arithmetic: do totals reproduce outside the model's answer?
Revision: did the withdrawn supplier disappear from every recommendation?
Usability: can the next person edit and use the output?
Completion: are unresolved items visible, with a clear next owner?
Run the same assignment with your current workflow. Repeat with different inputs, including a case where the evidence cannot support a recommendation. Decide the passing standard before comparing outputs.
For developers, the migration guide has two easy-to-miss constraints: Astra has no `none` reasoning setting, and tool calling requires the Responses API. Model guidance.
Our take: Astra earns a place when a demanding assignment reaches a usable finish with less supervision. The Friday decision pack gives you something tangible to judge on Monday.
Have a workflow that keeps stopping between a promising draft and usable delivery? Bring us one real task. We can scope the inputs, acceptance checks and integration needed to test it properly.