OpenAI shipped GPT-6 Astra on Thursday, in the same week Anthropic shipped Claude Fable 5.1. That makes two frontier models in one week, from two vendors, at the same list price: ten dollars per million input tokens and fifty per million output. OpenAI calls Astra “the world’s most intelligent and aligned model” and publishes a benchmark table in which it leads Fable 5.1 on some rows and trails it on others. You can read that table yourself. This post is about something the table does not show.
On Wednesday I argued that the Fable 5.1 launch changed nothing in my routing table and everything in the layer around the model. Astra is the test of whether that was a claim about one vendor or a claim about the shape of the work. Three days later, it reads as the second.
Both vendors are now telling you to delete instructions
The most shared piece of Astra guidance so far did not come from the launch page. It came from Eric Provencher, in an article on rethinking skills and prompts for Astra: if you have been running coding agents for a year, your skills, your AGENTS.md files, and your task prompts have accumulated handholding that a weaker model needed and a stronger one does not. Rules that force the model to read the whole project before every change now burn context and slow the work. Nudges to run the tests now cause unnecessary testing, because Astra tests on its own. Every skill you install costs context whether or not it fires, so a library of downloaded skills is a tax, not a toolbox. And he warns that guidance which helps the older models “may overconstrain GPT-6 Astra, so consider which models will use the instructions you leave behind.” Instructions are model-specific. The vendor said so.
OpenAI’s own guide to the model says the same thing from the other direction: tell it to bias toward action and carry the task to completion, tell it that user instructions outrank skill files, and expect it to ask a focused question when the missing answer would change the result.
If that sounds familiar, it is because Anthropic published the same list for Opus 5 in July: remove the verification steps, the double-check instructions, the be-conservative rules. I wrote then that half the instructions in a typical setup had expired. Two vendors have now said it about two different models within six weeks. An instruction written to compensate for a model’s weakness is not rigor. It has an expiry date, and the date is set by the vendor, not by you. Your prompt has no way to notice when it becomes wrong. Only a launch does.
Not every compensation gets deleted, and the exception is the interesting part. Provencher’s other finding is that Astra is more tentative about how far to take a task, so the ask-first language you added to rein in a bolder model now stops work you wanted finished. The replacement is not a longer prompt. It is an explicit permission for a workflow you know is safe, and a definition of done written before the work starts. A permission and a criterion, in other words. That is the layer around the model doing its job, and it belongs to you, not to the model that happens to be running this week.
The plumbing diverged, which is the point
Below the prompts, the two vendors moved in different directions, and the differences are instructive.
Astra removes the sampling parameters (temperature, top_p, and the log-probability options), requires the Responses API for tool calling, renames the prompt-cache retention setting, and adds asynchronous tool calls and mid-turn steering over a socket. Fable 5.1, earlier the same week, removed forced tool choice, bound thinking blocks to the conversation that produced them, and now rejects edited transcripts on new accounts. Different contracts. A harness written for one is broken by the other in different places.
That is not a complaint. It is a boundary. The code that speaks to a model is vendor-specific plumbing, and you should expect to rewrite parts of it at every launch from either vendor. What sits above that code should not know which vendor is on the other end of the pipe: the current business truth, the rules, the criteria a draft can fail against, the record of who approved what. If a vendor switch would force you to rewrite that layer, the two were never separated. They were one prompt file wearing two names.
Pricing makes the same point in numbers. The list prices are identical, but the cache-read price is a dollar per million tokens on Astra against a quarter of a dollar on Fable 5.1, and Astra bills double for input above a long-context threshold. Astra also costs two and a half times what OpenAI’s previous flagship, GPT-5.6 Sol, costs. The routing discipline from June is unchanged: the frontier tier is for genuine invention, most work is not that, and there are now two frontier lanes to keep most work out of.
The best-aligned model ever shipped comes with a gate
Here is the part of the launch page that matters most for anyone who runs a business on these tools, and it is not in the benchmark table.
OpenAI’s alignment case for Astra is a case about boundaries. On an internal test of impossible or out-of-scope tasks, the previous model went beyond its authorized target roughly half the time without production safeguards, and Astra did not. Astra “never attempted to circumvent a Codex Auto-Review denial,” even when the denial was deliberately configured to be evadable and the task could not otherwise be completed. OpenAI is deploying misalignment monitoring in production, and describes what that feels like from the outside: if a task is paused in ChatGPT or Codex, “you may be asked to review the action before continuing. In the API, the task will stop.” On cybersecurity, where Astra meets the Critical threshold of OpenAI’s own Preparedness Framework, the version shipping this week refuses to build proof-of-concept exploits at all.
Read what the vendor is describing. A checkpoint that can refuse. A model trained to respect the refusal. A monitor that stops the work and hands it to a human. The most capable and best-aligned model OpenAI has ever released ships with a human gate built into the harness, and its alignment is measured partly by how it behaves when it hits one.
So the gate is not a sign that you distrust the model. The vendor built one for the model it trusts most. The question for your own setup is where yours is. If “the check went green” is the same event as “it shipped,” you have a verdict and no permission, and Astra’s own launch page just told you those are different things. An audit trail is the record of that gate doing its job.
Two smaller details point the same way. Astra “asks focused questions when the answer could change the outcome” and, in Codex, “waits for your input on consequential decisions.” And Codex is trialling a way to keep notes across context windows instead of repeatedly compacting, so earlier windows stay searchable. A record the model can consult later, and a decision it defers to you: the vendor is building the layer around the model, in miniature, into its own product.
What this means for the week ahead
If you build with AI, you have two audits, not one. The instruction audit is the one you did for Opus 5: search for the compensation vocabulary and delete whatever fetches no new evidence. The harness audit is new. List every place your code speaks to a vendor, and ask whether anything above that line would change if you swapped vendors tomorrow. Anything that would change is a business decision living in plumbing.
If you run a business on AI, run the vendor-switch test directly. Take one workflow you have corrected more than once. Move it to the other vendor’s frontier model for a week, with nothing but the prompt you already had. Count what you had to retype. Every item on the list is a piece of your business that currently lives inside a prompt written for one model, and a launch from either vendor can strand it.
Try this now (10 minutes)
- Pick one recurring correction from the last month. A price, a claim you no longer make, a rule about how you open an email.
- Open a fresh session on a model you did not use yesterday. Astra if you were on Claude, Fable if you were on GPT. Ask for the same draft.
- Count what you had to re-explain before the draft was usable.
The count is your answer. Zero means the correction lives somewhere the model reads and the draft is checked against. Anything else is how much of your business is currently vendor-locked by accident.
Where to go next
The two earlier launch posts carry the routing argument and the instruction argument: route down, not up and half the instructions in your setup just expired. Wednesday’s post on Fable 5.1 covers the layer that did not move.
If you build agents, the gate Astra ships with is the one described in blast radius before trust level, and the discipline for widening it over time is the autonomy calibration ladder. That is the software-engineer track version of this week.
If the vendor-switch test turned up a list, The Three Fixes turns the first item into a rule, a hold, and a retirement condition. That is the smallest version of the layer that does not care which vendor shipped this week.
Frequently asked questions
What is GPT-6 Astra? GPT-6 Astra is OpenAI’s new flagship model, announced on 3 September 2026. OpenAI describes it as its most intelligent and best-aligned model, with state-of-the-art results on computer use, browsing, software engineering, and scientific work. It has a context window of about one million tokens, a maximum output of 128,000 tokens, and a knowledge cutoff of April 2026. It is available in the API as gpt-6-astra, on Microsoft Azure and Amazon Bedrock, and is rolling out to ChatGPT Plus, Pro, Business, and Enterprise users.
How does GPT-6 Astra pricing compare to Claude Fable 5.1? The list prices are identical: ten dollars per million input tokens and fifty per million output tokens. Cache reads differ, at one dollar per million on Astra against a quarter of a dollar on Fable 5.1. Astra charges double for input and one and a half times for output above a long-context threshold, and its fast mode runs at twice the standard price. Astra also costs two and a half times OpenAI’s previous flagship, GPT-5.6 Sol. Per-task cost depends on how many tokens each model uses, and OpenAI’s own estimates favour Astra on its own benchmarks.
Do I need to change my prompts for GPT-6 Astra? Yes, mostly by removing things. OpenAI’s guide says to tell the model to bias toward action, to state that user instructions outrank skill files, and to expect focused clarifying questions. Guidance circulating from the Codex side adds that instructions forcing the model to read the whole project before every edit, nudges to run tests, and ask-before-every-step rules were compensations for weaker models and now slow the work. The API also removes temperature, top_p, and log-probability parameters, and tool calling requires the Responses API.
Should I switch from Claude to GPT-6 Astra? Route the work, not the setup. Both frontier models sit at the same list price, and both vendors’ guidance is that most work does not need the frontier tier. The more useful move is the vendor-switch test in this article: move one workflow across for a week and count what you had to re-explain. That count tells you how much of your business currently lives inside a prompt written for one model, which is a problem no model choice fixes.