Two top models in one week: Anthropic released Claude Fable 5.1 on 1 September, OpenAI released GPT-6 Astra on 4 September — the model people are currently searching for as “ChatGPT Astra”. Since then the same question has been reaching me in different forms: ChatGPT Astra or Claude — which is better now? Here is the answer, as honest as I can give it — with numbers that are documented, and without duel drama.
The short version first: both cost the same pay-as-you-go, both have a million tokens of context, both ship with a brake on. The difference lies elsewhere: in the plan, in continuous operation, and in what you intend to do with it.
ChatGPT Astra or Claude: what you get on a plan
This is the point that matters most to you as a user — and the one usually missing from the duel articles. Because the two makers handle their top model on a plan differently:
| Plan | Price per month | Top model included? |
|---|---|---|
| ChatGPT Plus | €23 incl. VAT | yes — GPT-6 Astra |
| ChatGPT Pro | €229 incl. VAT | yes — GPT-6 Astra Pro |
| Claude Pro | €15 net on the annual plan | only via extra usage credits; on the plan: Sonnet and Opus 5 |
| Claude Max | from €90 net | yes — Fable 5.1, up to half of the weekly limits |
| Free (both) | €0 | no |
Translated: if you have ChatGPT Plus, you simply get the top model for 23 euros. With Claude you need Max for that — or you work on the Pro plan with Opus 5, which Anthropic recommends for most tasks anyway. Fable 5.1 is the model there for tasks that should run for hours without you, not the model for everyday text.
I checked the plan prices in August; the Claude plans in detail are in What does Claude cost?. The makers change their packages more often than one would like — when in doubt, check directly with the provider.
What the models cost pay-as-you-go
If you are not working on a plan but are billed by consumption — say because an agent runs for you or you built something yourself — the bill looks like this:
| Price per M tokens | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|
| Input | $10 | $10 |
| Output | $50 | $50 |
| Cache read | $1 | $0.25 |
| Fast mode | double the price, up to 2.5× faster | — |
| Context window | around 1M tokens | 1M tokens |
The base prices are identical. The difference is in re-reading: when an AI works on a task for a long time, it reads the same material over and over — your files, the history, the instructions. That runs through the cache, and the cache is four times cheaper with Claude. Anthropic calculates around 25 percent lower total costs for typical use, and up to 45 percent for tasks that run on their own. The background: Claude Fable 5.1: stronger than Opus 5 and cheaper to run.
Astra in turn has something Claude does not offer: a fast mode that answers up to two and a half times faster at double the price. Who that matters for: everyone whose waiting time costs more than the bill.
Where each model leads
GPT-6 Astra: operating the computer. That is the number OpenAI rightly puts up front. In the OSWorld 2.0 test — real tasks on screen, opening programs, clicking, editing files — Astra completes 72.6 percent, its predecessor 65.7. It needs around 47 percent less time and finds the right button in crowded interfaces in 92.7 percent of cases instead of 76.9. In OpenAI’s own task collection it solves 88 percent on the first attempt, its predecessor 56. What that means for you and what still goes wrong is in ChatGPT Astra (GPT-6): what it costs, who gets it, what changes.
Claude Fable 5.1: long chains, cheap and concise. Fable 5.1 beats Opus 5 in all comparison tests Anthropic has published, most clearly in scientific work, where the score has more than doubled compared with Fable 5. Add the reworked writing style — shorter, less preamble — and the cheaper continuous operation from above.
An honest note on all of these numbers: each comes from the maker. An independent test that puts both models directly against each other is not something I can cite as of today. Each measures what it is good at. That is why I consider the test scores the smaller part of the decision.
What is the same for both
- A million tokens of context — both read entire folders in one go.
- The same base prices — $10 in, $50 out per million tokens.
- Shipped with a brake on. Astra is the first OpenAI model in the highest internal risk level for cybersecurity; you get a restricted version that refuses certain requests. At Anthropic the same pattern is called Fable 5.1 for everyone, Mythos 5.1 only for vetted organizations. For your daily work both are irrelevant.
- Both get better every few weeks — without you doing anything.
My experience — and where it ends
So you can place my recommendation: 90 percent of my business runs on Claude, on the Max plan — websites, apps, automations, analyses, the newsletter. I used ChatGPT on the highest tier for half a year in 2025 and got considerably less out of it for myself. The three reasons for that are in Claude vs. ChatGPT: which AI fits your business?.
Where my experience ends: I have not yet run GPT-6 Astra in my own daily work. What I know about it comes from the announcement and the published figures. So what I am writing here is not a hands-on report on Astra but an assessment. As soon as I have tested it myself, I will add that.
And one point for ChatGPT that still holds: custom assistants you hand to clients work there as custom GPTs — Claude has no comparable route for that so far. If that is your business model, it decides more than any test score.
Who should switch — and who should not
- You are on ChatGPT Plus and happy: stay. Astra turns up in your model menu on its own, and for 23 euros you have the top model included. Switching to Claude gives you nothing now that you do not have — except the effort of rebuilding your projects and instructions.
- You are on Claude: stay. Your daily work runs on Sonnet and Opus, you reach for Fable 5.1 only with good reason — and that reason has just become cheaper. That Astra operates the computer better is something Anthropic is catching up on in the same discipline right now; for you it is no reason to move your workflows.
- You are starting fresh: do not decide by this table, but after a week with a real project. Both have a free tier that is enough for that. Two questions help: do you want to build — pages, apps, systems? Then look at Claude Code, that is what mine were built with. Do you want to hand assistants to clients? Then ChatGPT.
And the sentence that stands above everything since I read both models side by side: the difference between the top models is smaller than the difference between a tool that lives in your workflows and one you are just getting to know. Whoever switches loses the time the new model was supposed to save.
Nobody knows which model will be ahead in four weeks. Whether your workflows benefit is up to you. That is exactly what we talk about in my free community “KI — aber richtig” — the templates and workflows from my videos live there too. And if you want company instead of doing it alone: the AI Business Community offers monthly live calls and workshops for exactly that.
Frequently asked questions
ChatGPT Astra or Claude — which is better?
It depends on what you do with it. GPT-6 Astra leads at operating computers, Claude Fable 5.1 at cheap continuous operation and at long tasks that run on their own. Both figures come from the respective maker; there is no independent head-to-head yet. For most self-employed people the deciding factor is where projects, instructions and routines already live.
What does ChatGPT Astra cost compared with Claude?
Pay-as-you-go both cost the same: $10 per million input tokens and $50 for output. Cache reads cost $1 with Astra and $0.25 with Claude. On a plan, Astra is included in ChatGPT Plus for €23 incl. VAT; with Claude, Fable 5.1 is on the Max plan from €90 net, and on the Pro plan only via extra usage credits.
Is GPT-6 Astra included in ChatGPT Plus — and Fable 5.1 in Claude Pro?
Astra yes: since 4 September 2026 all Plus users have been getting the model enabled in waves. Fable 5.1 no: on the Claude Pro plan it only runs via extra usage credits; it is included in Max, Team Premium and Enterprise Premium.
Is it worth switching from Claude to ChatGPT Astra?
For most people, no. Whoever has built projects, instructions and routines in Claude loses more time moving than GPT-6 saves. The same applies the other way round to ChatGPT users who want to switch to Claude. A switch only pays off if you need something your tool cannot do — such as handing custom assistants to clients.
Conclusion
ChatGPT Astra or Claude Fable 5.1 is not a question a table answers. Pay-as-you-go both cost the same, both read a million tokens, both run with a brake on. On a plan, Astra is the cheaper route to a top model; in continuous operation, Claude is cheaper. Whoever already works with one of them stays. Whoever is starting fresh tests for a week with a real project — and decides after that, not by headlines.
