A closer look at its new capabilities, the Fable 5.1 competition and what could change in everyday work.
OpenAI launched GPT-6 Astra on September 3, 2026, beginning with limited organizational access before a broader rollout. Its latest model brings improvements across coding, computer use, research and document creation.
What interests me about Astra is how those capabilities might work together. A model that can research a subject, organize the findings and produce a usable document could be valuable even when I am not writing code. The question is how much correcting and supervising that process still needs.
What GPT-6 Astra Brings to Everyday Work
Astra is designed to carry work across browsers, code and professional software. It can incorporate new requirements, answer side questions and ask for information that could change the outcome. OpenAI also warns that its tendency to ask questions can interrupt work when a user expects it to make a reasonable assumption.
That distinction matters to me. If I ask for a report, I want a sensible question about the audience. I do not want to approve every formatting choice. Better judgment about when to continue and when to ask would make the assistant more useful than simply giving it more tools.
The release also targets documents, spreadsheets and presentations that follow supplied templates. For productivity work, I would look closely at whether Astra preserves the original structure, carries the right facts into it and leaves something I can actually edit.
Why Computer Use Is the Part to Watch
OpenAI reports 72.6% on its offline, partial-scoring OSWorld 2.0 evaluation, versus Sol’s 65.7%. That gives me a reason to expect better performance in software interfaces. It does not tell me whether Astra can reliably finish a messy workflow involving several applications.
The test I would care about is a complete piece of work: collect information from several pages, put it into a spreadsheet, check the calculations and turn it into a short presentation. If a source changes or a browser tab fails halfway through, can Astra recover without losing the assignment?
Astra Versus Claude Fable 5.1
Claude Fable 5.1 is the comparison I would put beside this launch. Anthropic released it on September 1 for demanding coding and long-running work, making it a current alternative for the same kinds of assignments.
The graph below compares selected published results without combining them into a winner score. It includes an area where Fable leads as well as areas favoring Astra.
I have deliberately left OSWorld out of this head-to-head graph. Astra’s cited result uses an offline subset, while Anthropic reports a differently configured evaluation. Putting the two percentages side by side would invite a comparison the setups do not support.
Independent testing gives Fable a stronger counterpoint. Artificial Analysis reports Coding Agent Index scores of 70 for Fable 5.1 and 67 for Astra. Astra uses about one-third of Sol’s tokens in that coding evaluation, but its broader Intelligence Index result is approximately level with Sol.
For an established Fable workflow, my first Astra trial would be the assignment that currently takes the most correction. Changing models should solve a recognizable problem, and the finished work is where that difference needs to appear.
Availability and Pricing
The announced rollout includes ChatGPT Plus, Pro, Business and Enterprise, alongside the API, Azure and AWS Bedrock. Pro, Business and Enterprise are also slated for Astra Pro. Enterprise access is initially off by default. OpenAI still describes access as limited rather than generally available, so plan eligibility does not guarantee that the model is visible in every account. Usage is included within existing subscription allowances, with additional credits planned for extra usage. API rate limits vary by usage tier.
Astra’s standard API rates are $10 per million input tokens and $50 per million output tokens. Cache reads cost $1 per million tokens. Fable 5.1 has the same headline input and output prices, but cache reads cost $0.25.
That cache difference is worth noticing if an application repeatedly sends the same large context. It does not settle the total bill: the number of tokens used, retries and processing mode still matter.
Artificial Analysis found Astra’s maximum-effort coding cost approximately matched Sol’s while producing a better score. In its broader Intelligence Index evaluation, Astra was roughly 75% more expensive per task. The value therefore depends on the assignment, not just the price per token.
Science and Reasoning Broaden the Release
The scope extends beyond office work. The launch includes mathematics, scientific workflows, biology and health evaluations. A strong score is a reason to investigate a research assistant’s usefulness, not a substitute for checking its calculations or conclusions.
ARC Prize reports 99.9% on ARC-AGI-3 with a provider adapter, versus 62.7% with its standard harness. The adapter preserves internal reasoning between requests and compacts conversations; the best scores also use different reasoning settings. This measures a complete setup, not what every application will reproduce.
What Developers Should Know Before Switching
Astra accepts text and images and produces text, with a 1.05-million-token context window and up to 128,000 output tokens. Long inputs have a pricing catch: above 272,000 input tokens, the entire request carries higher rates.
For applications, asynchronous tool calls allow other work to continue while a tool runs; mid-turn steering allows a correction before the response has finished. Tool calling requires the Responses API. Astra also drops support for the none reasoning level and custom temperature or top_p values. These are useful capabilities and compatibility checks, not features that every existing app automatically gains.
OpenAI’s guidance flags another practical issue: Astra can perform broader testing than a small coding task requires. I would want it to verify a change properly without turning a minor fix into a much larger assignment.
The Safeguards Are Part of the Experience
OpenAI places Astra at its Critical cybersecurity capability threshold and applies additional safeguards. That is a risk classification, not a reliability certificate. The system card also acknowledges limitations in monitoring its reasoning.
Safety monitoring can pause or stop work for review. I would watch how frequently legitimate tasks are interrupted and whether the explanation makes the next step clear. An assistant that operates software needs boundaries, but users also need to understand why a task did not finish.
My Take on the Astra Launch
Astra gives me a reason to try more demanding development and productivity tasks. Better computer use could make it useful inside the software where the work already happens, but complex workflows will be the more convincing test.
Can it handle a changed brief, catch a mistake and deliver something usable without constant intervention? That is what I want to find out, with Fable 5.1 alongside it. A clear improvement there would matter more to me than another overall ranking.