In June 2026, Anthropic released a new flagship model: Claude Fable 5. The official documentation lists it as Anthropic’s “most capable broadly available model”. Since then I have been getting the question that comes with every new model: Do I have to switch now?
The honest answer up front: It depends on what you actually do with it — and the gap between the two answers is bigger than I would have thought myself. For part of my work, Fable 5 has opened a completely new door. For the rest it would be burning money, because it is not even included in your Pro plan.
Here is the sober comparison of the three models — and where the line runs.
What Claude Fable 5 actually is
Fable 5 has been generally available since 9 June 2026. Anthropic positions it as a model for long-running agents — that is, for tasks that run for hours and must not lose the thread, not for the quick chat in between.
The numbers Anthropic advertises are impressive: at Stripe, the model completed a migration of a Ruby codebase in one day that had been estimated at two months. On analysis tasks it scores ten points above its predecessor. On tasks with persistent memory, three times as much.
Those are real leaps. But: they are leaps in disciplines that have little to do with the everyday life of a small business. A Ruby migration is not a newsletter.
The story you should know
Part of the context is that after launch, Fable 5 was completely shut down for three weeks.
On 12 June, the US government applied export controls to the model — triggered by a report from Amazon researchers who managed to bypass its safety mechanisms. Because the order took effect immediately and users’ nationality cannot be checked in real time, Anthropic blocked access for everyone. On 30 June the controls were lifted; since 1 July the model has been back worldwide.
I am not telling you this to scare you — but because there is a lesson in it that holds independently of this one case: Never build your processes so that they depend on a single model. A tool that can be gone overnight must not be the sole carrier of your day-to-day business. That goes for Fable 5 just as much as for any other tool that feels indispensable to you right now.
The comparison: Fable 5, Opus 4.8 and Opus 4.6
All figures come from Anthropic’s official model documentation, as of July 2026.
| Fable 5 | Opus 4.8 | Opus 4.6 | |
|---|---|---|---|
| Intended for | long-running agents | complex work, everyday tasks | previous generation |
| Price (API, per million tokens) | $10 in / $50 out | $5 / $25 | $5 / $25 |
| Context window | 1 million tokens | 1 million tokens | 1 million tokens |
| … that is roughly | 555,000 words | 555,000 words | 750,000 words |
| Speed | slower | medium | medium |
| Thinking mode | always on | controllable | controllable |
| Knowledge cutoff | January 2026 | January 2026 | May 2025 |
| Status | current | current | Legacy |
The three differences that really matter day to day
1. The price — and what your plan includes
In the API, Fable 5 costs twice as much as Opus 4.8: 10 instead of 5 dollars per million input tokens, 50 instead of 25 for output.
Far more important for you, though, is what happens in your subscription. On Pro, Max and Team, Fable 5 was part of the normal usage limits only until 7 July. Since then, access runs on usage credits — meaning you pay per use on top of your subscription. The free plan does not include the model at all.
This is the point where the question settles itself for most people: The strongest model is not the one that comes with your plan. Opus 4.8, by contrast, is included in Pro and Max as standard. I have broken down what the individual plans look like for you here: What does Claude cost?
2. The speed — and the fact that you cannot turn it off
In Anthropic’s own overview, Fable 5 is rated slower, Opus 4.8 as medium. That is not a flaw but by design: on Fable 5, the thinking mode is permanently active and cannot be switched off. The model always thinks things through thoroughly — even when all you need is a quick subject line.
For an agent working on a task for eight hours, that is exactly right. For the twenty small questions you ask on a normal working day, you pay with waiting time for a thoroughness you do not need.
3. The knowledge cutoff — the real reason to leave 4.6 behind
Here lies the difference hardly anyone has on their radar. Fable 5 and Opus 4.8 both have a knowledge cutoff of January 2026. Opus 4.6, however: May 2025.
That is eight months’ difference — and in this field, eight months is an eternity. A model with a May 2025 cutoff simply does not know what has happened since. It cannot tell you that; it just answers from its old state of knowledge. Opus 4.6 is by now also officially filed in the documentation under Legacy.
A second point that becomes practical when you work with long documents: since generation 4.7, Claude uses a different way of splitting text into tokens. The same text therefore produces around 30% more tokens than before. All three models do have a context window of one million tokens — but for Opus 4.6 that equals roughly 750,000 words, for Opus 4.8 and Fable 5 only around 555,000. The same window, less text inside.
What has changed for me personally
So much for the spec sheets. Now the part you cannot read from any table — my own experience with the two models in everyday work.
For me, Fable 5 has opened up an entirely different dimension. I have clearly experienced the differences in results between Opus 4.8 and Fable 5 in four areas:
- Building websites
- Agent mode — that is, tasks the AI works through independently across several steps
- Running analyses
- Turning an idea into an app
For me this really is a completely new world. And I only realised how much when the model was gone: The sudden shutdown set me back completely. Not a little slower — set back. After Fable 5 returned, I doubled my personal Max subscription.
Why I still will not sell you this as a recommendation: look at the four points above again. Those are all tasks where the AI works long and independently — exactly what Fable 5 is built for. If your day-to-day looks like that, the surcharge is worth every cent. If you mainly write texts, sort ideas and plan content, you simply will not feel the difference — and you pay for capabilities you never call on.
Incidentally, that is also the answer to why I am not giving a general buying recommendation here: The model question is not a tool question — it is a process question. What is right for me follows from what I build — not from it being the strongest model. And what is right for you follows from your business, not from mine.
And what about Opus 5?
Since this comparison, another model has arrived: Opus 5. If you are wondering whether to take Fable 5 or Opus 5 now — here is the shortest honest answer:
Fable 5 is the strongest model. Opus 5 is the one most people are better off with.
Opus 5 is a clear leap over Opus 4.8 and particularly strong at everything that runs across several steps: longer tasks, work with many files, research that is not finished after one answer. And it costs half of Fable 5 — 5 instead of 10 dollars per million tokens in, 25 instead of 50 out.
The decisive point for your subscription: Opus 5 costs the same as Opus 4.8 and is still the stronger model. That leaves no reason to stay on 4.8 — just as there was no reason left to stay on 4.6 before.
- Opus 5 — 5 dollars in, 25 out. One million tokens of context. Included in the Pro plan (as of July 2026).
- Fable 5 — 10 dollars in, 50 out. Likewise one million tokens of context. Included in the Max plan; on Pro, access runs on additional usage credit (as of July 2026).
Both have the same context window. In terms of how much text you can process in one go, the two are equal — the difference lies solely in the depth of the work, and in the price.
Which one do you take? For everything that comes up day to day — texts, research, evaluations, building pages, automating workflows — Opus 5 is enough, and you pay nothing extra. You reach for Fable 5 when a single task is genuinely hard and runs long: a big migration, research across many sources, something that keeps working for hours without you.
It is the same logic as above: the strongest model does not win — the one whose strength fits your task does.
And what about Sonnet 5?
While we are on the question of which model you need, it would be dishonest to talk only about the expensive top end. Because for many, the answer is not at the top at all.
Since 30 June 2026 there has been Sonnet 5 — and it is the model talked about least, even though it is the most obvious one for most small businesses.
Sonnet 5 is not the stripped-down variant. It is the fast one.
Anthropic describes it as the best combination of speed and intelligence, with quality close to Opus on coding and agent tasks. The context window is the same as Opus 5 and Fable 5 — one million tokens, knowledge cutoff January 2026. In the API it costs a fraction of that.
| Model | Input | Output | Context window |
|---|---|---|---|
| Sonnet 5 | $3 ($2 until 31 Aug) | $15 ($10 until 31 Aug) | 1 million tokens |
| Opus 5 | $5 | $25 | 1 million tokens |
| Fable 5 | $10 | $50 | 1 million tokens |
Per million tokens, net. Until 31 August 2026 an introductory price applies to Sonnet 5; after that it costs a good half of Opus 5 and just under a third of Fable 5.
And now the part hardly anyone adds: These prices apply to the API — that is, when you pay per request because you built something yourself. If you use Claude on a subscription, you pay a fixed monthly amount no matter which model you click. So in the subscription, Sonnet 5’s price advantage is no advantage at all.
What remains in the subscription is something else: Sonnet 5 answers faster. Everywhere you wait for the answer and then keep working — drafting texts, sorting ideas, evaluating a transcript, outlining a post — you notice that every day. With Opus you wait longer for a result that is often no better at this point.
Which one do you take?
- Sonnet 5 for everything you need often and fast: writing, rephrasing, sorting, summarising, evaluating.
- Opus 5 for everything where a thinking error gets expensive or where work runs across several steps: analyses, structure, building pages, automating workflows.
- Fable 5, when a single task is genuinely hard and keeps running for hours without you.
The honest order for most people is therefore: Sonnet for everyday work, Opus for the heavy lifting, Fable only with good reason. And you do not have to commit — you switch per task. Try Sonnet 5 for a week on your daily text work before you assume you need Opus.
And what about Haiku?
Haiku 4.5 is the fastest model in the line-up: 1 dollar in, 5 dollars out per million tokens. In return, a smaller context window of 200,000 tokens and a knowledge cutoff of February 2025 — considerably older than all the others. For short, simple and very frequent tasks it is the cheapest choice. Wherever up-to-date knowledge matters, the old cutoff is the knockout criterion.
Where you actually switch the model
Choosing a model sounds more technical than it is — you make the choice in a place you already know. In Claude you pick the model right in the chat, via the model picker above or below the input field. You have the same choice in Claude Code and in Claude Cowork.
Two practical notes on that:
- You do not have to decide once and for all. You can switch per task — and that is exactly the sensible way to handle it. Choosing one model for your whole business is about as clever as buying a single screwdriver.
- One special rule applies to Fable 5: Anthropic retains inputs and outputs for this model for 30 days for safety review. You should know that when deciding what to work with.
And if you have only used Claude in the browser so far: the model picker works exactly the same in the app — on your computer and on your phone. Where to get it and what your device needs is covered in Claude app and download.
So: which model do you take?
If I sort them by the everyday life of a small business, it looks like this:
- Opus 5 is the answer for almost everyone. It is included in the subscription, built for complex and long work, and costs the same as Opus 4.8 despite being stronger. For offers, newsletters, analyses, content and structure it is more than enough.
- Fable 5 pays off when you really run long chains — building websites, working in agent mode, running analyses, turning an idea into an app. Tasks where the AI works independently for hours and a mistake in step three breaks step seventeen for you. That is exactly where I experienced the difference — and why I doubled my Max subscription. For everything else you pay for capabilities you never call on.
- Sonnet 5 is the underrated candidate. Faster than Opus, with the same knowledge cutoff and the same context window — and on a subscription, trying it costs you nothing. For the bulk of daily text work it is completely sufficient. Try it before you assume you need Opus.
- Opus 4.8 and 4.6 are no longer an argument. All three Opus tiers cost the same, but Opus 5 is the strongest of them. If you are still set to 4.6 or 4.8 anywhere, this is the one switch that pays off immediately — it costs nothing and delivers right away.
And that brings us to what I keep saying here: The most expensive model is not automatically the right one. The question is never “What is the best?” but “What is good enough for what I plan to do with it?” You can only answer that question if you know your processes — not if you know the benchmarks.
If you first want to understand what Claude can actually do for your business, start here: What is Claude AI? The guide for solopreneurs.
Choosing the model is the smaller part. The bigger one is guiding it so something usable comes out. 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, workshops and the MACHZEIT! challenge for exactly that.
Frequently asked questions
What is the difference between Opus 5 and Fable 5?
Fable 5 is the stronger model, Opus 5 the better fit for most. Opus 5 costs half — 5 instead of 10 dollars per million tokens in, 25 instead of 50 out — and is included as standard in the Pro and Max plans. Fable 5 is included in Max; on Pro only via additional usage credit. The context window is the same size for both (one million tokens). The difference becomes noticeable mainly on tasks that run long and independently.
What is Claude Sonnet 5 good for?
For everything that should be quick and comes up often: drafting and rewriting texts, evaluating transcripts, sorting ideas, outlining posts, summaries. Sonnet 5 is close to Opus in quality but answers faster and costs around half in the API. For tasks where work runs across several steps or a thinking error gets expensive, Opus 5 is the better choice.
Is Sonnet 5 available in Claude Code?
Yes. In Claude Code you choose the model just as in the chat — Sonnet 5, Opus 5 and Fable 5 are available there, depending on your plan. For short, frequent steps Sonnet is often the more pleasant choice because it answers faster.
Is Claude Fable 5 included in the Pro plan?
No — but it is in Max. Since 7 July 2026, Fable 5 access on the Pro plan runs on usage credit billed on top of the subscription. In the Max plan the model remains included as standard (as of July 2026). The free plan does not offer it. Opus 5, by contrast, is included as standard in Pro and Max — and the better fit for almost everything.
Is Claude Fable 5 better than Opus 4.8?
On the measured tasks, yes — Anthropic reports clear leads, around ten points on analysis tasks. In practice it depends on what you do: for building websites, working in agent mode, analyses and turning an idea into an app, the difference is clearly noticeable. For a newsletter or an offer page you will hardly notice it — and Fable 5 is twice as expensive, slower and only available on the Pro plan via additional usage credit.
Why was Claude Fable 5 temporarily unavailable?
The US government applied export controls to Fable 5 on 12 June 2026, after Amazon researchers had found a way to bypass the model’s safety mechanisms. As the order took effect immediately, Anthropic blocked access for all users. On 30 June the controls were lifted; since 1 July 2026 the model has been available worldwide again.
Should I still use Claude Opus 4.6?
Probably not. Opus 4.6 costs the same as Opus 4.8 but has a knowledge cutoff of May 2025 instead of January 2026 and is listed as legacy in Anthropic’s documentation. There is currently no argument for 4.6 and against 4.8.
How much text fits in a context window of one million tokens?
That depends on the model, even though the number looks the same. Since generation 4.7, Claude splits text into tokens differently — the same text produces around 30% more of them. For Opus 4.6, one million tokens equals about 750,000 words; for Opus 4.8 and Fable 5 only around 555,000.
Conclusion
Fable 5 is an impressive model — and whether it is the right one for you depends not on how good it is, but on what you plan to do with it. For me it opened a new door, because I build things that run long and independently. If you mainly write and plan, you will not experience the same difference — and you pay extra for it.
So the honest recommendation looks like this: Take Opus 5, as long as you have no concrete reason for anything else — it is included in the subscription, costs the same as the older Opus tiers and is strong enough for almost everything. Try Sonnet 5 for your daily text work before you assume you need Opus. Look at Fable 5, once you start working with agents or building things of your own. And if you are still stuck on 4.6 or 4.8 somewhere: that is the switch worth making today.
And if you would rather not figure this out alone: That is exactly where my Claude Sprint comes in. Not at the question of which model is currently ahead — but at your processes, from which the answer emerges in the first place. Five days live, in a small group, and you implement Claude directly in your business.
