The most powerful Claude model Anthropic has ever made available to the general public just landed in the subscriptions — and it's already known that on 23 June it will disappear from them. This isn't your typical launch, where a new model simply stays on the price list. This time you get a time window, and after it the rules change. In this article I'll explain exactly what Anthropic shipped, how Fable 5 differs from Mythos 5, what it costs and what specifically to do over the next two weeks.
Two models, one launch
Anthropic announced two models at once: Claude Fable 5, available to everyone as of today, and Claude Mythos 5, reserved exclusively for partners in the Project Glasswing program.
To understand why this matters, you need one term: the Mythos class. That's what Anthropic calls the layer of Claude models that, in terms of capability, sit above Opus — until now the most powerful model in the lineup. The first member of this class was Claude Mythos preview, released in April through Project Glasswing — the company's collaboration program with a narrow group of trusted parties. Now two more join: Fable 5 and Mythos 5.
The most interesting line from the announcement runs roughly like this: Fable 5 is a "Mythos-class model that we've made safe for general use." In other words — for the first time, every user is handed a model from a shelf that until now had been locked.
The same engine, different safeguards
Here's the heart of the whole launch: Mythos 5 is exactly the same model as Fable 5. The difference is that in Mythos 5 the cybersecurity safeguards have been removed.
What does that mean in practice? Safeguards are restrictions built into the model — rules that make it refuse to help with tasks deemed risky, for example breaking into other people's systems. Fable 5 has those restrictions switched on, which is why it can reach everyone. Mythos 5 has them removed, so it goes only to a narrow circle of cyberspace defenders and critical-infrastructure operators — deployed through Project Glasswing, in collaboration with the US government, as the successor to the preview version. Anthropic describes it as the model with the strongest cybersecurity capabilities in the world. The Glasswing program is meant to expand over time into broader "trusted access," but there's no sign it will reach ordinary consumers.
In my view, this design says something important about the very nature of models. At their core they're neutral — the same capability that lets them write code brilliantly also lets them break that code. The direction is decided by the intent and the judgment of the person who runs the model. Anthropic isn't pretending the risk doesn't exist; instead it has split one technology into two versions, with different levels of trust in the recipient. Whatever you think of the company itself, that's a more honest arrangement than pretending the problem will solve itself.
What it costs — tokens in plain terms
On the API both models cost the same: $10 per million input tokens and $50 per million output tokens.
A quick explanation before those numbers tell you anything. A token is the unit in which models "count" text — roughly a fragment of a word; a thousand tokens is about 750 words. Input tokens are everything you send to the model (the question, documents, context); output tokens are everything the model writes back. You pay separately for each.
Those rates mean twice the price of Opus — a model that already had a reputation for being expensive. For comparison: Claude Mythos preview cost Glasswing partners five times Opus, so the direction is downward rather than upward. And honestly: if the model's capabilities are what they appear to be, twice Opus for a jump of a whole class higher isn't a prohibitive price in the scale of real work. An expensive model that cuts a task from a day to an hour can be cheaper than a cheap one that can't.
The window until 22 June — and what happens next
Now the part that calls for a decision from you, not just reading.
From today until 22 June, Fable 5 is part of the Pro, Max, Team and Enterprise plans (in the per-seat variant) — at no extra charge, within your regular subscription. On 23 June, Anthropic removes Fable 5 from those plans. From that day, using the model will require usage credits — a prepaid pool billed per token, separate from the subscription. The subscription stops being enough; you start paying for actual usage.
Anthropic states that it intends to bring Fable 5 back into the subscriptions "when compute allows" — and that it wants to do so as soon as possible. It doesn't give a date, though. Why the maneuver? The context is fairly clear: the company has just filed to go public, isn't profitable, and faces enormous compute costs. Offering the most expensive model in the company's history at the price of a regular subscription simply doesn't add up — at least not with the current infrastructure. That's understandable, but for you it means one thing: you have about two weeks of real, surcharge-free access.
What to do with it? Don't put off the testing "for a calmer moment." Pick two or three of your real, recurring tasks — the report you write every week, document analysis, a chunk of work on a website or a proposal — and run them through Fable 5 alongside the model you use today. After two weeks you'll know from your own practice, not from other people's enthusiasm, whether this model is worth the credits once the window closes. One technical note: if you don't see Fable 5 in the model list, update the app or tool you're using — older versions won't show it.
Benchmarks — what they show and how to read them
A benchmark is a standardized comparison test: the same set of tasks solved by different models, so they can be lined up on a single scale. According to Anthropic, Fable 5 achieves the best results on the market in almost every area tested.
I treat such claims with reserve — a maker praising its own product isn't an independent source — but the directions of change alone say a lot. Clear jumps show up in knowledge work (analysis, documents, research), in agentic programming, in legal tasks, in reasoning that spans multiple fields, in biology and in image understanding. The reference points in the comparisons are Opus 4.8 and GPT 5.5, the current front-runners. The jump in the "Frontier code" test, which measures hard programming tasks, especially stands out. One caveat: skip the SWE Bench Pro result — serious doubts have recently been raised about the credibility of that particular benchmark, and it's not worth building anything on it.
The most practical reference point concerns effort levels — the setting with which you control how long and how deeply the model "thinks" about a task, which translates directly into cost. According to the published data, Fable 5 at the low effort level achieves results close to Opus 4.8 at x-high — and comes out a bit cheaper in the process. And the maximum effort level in Fable 5 brings a noticeable improvement over x-high, something that was essentially invisible in the previous generation. So if today you squeeze everything out of Opus 4.8 on the highest settings, Fable 5 gives you a similar ceiling already in low gear — with headroom still left above it.
Agent loops — enthusiasm with a brake
The new models can work autonomously far longer than previous versions of Claude. On that wave, a fashionable view is gaining ground: stop writing instructions to agents, let agent loops drive other agents. An agentic loop is a setup in which the model works round and round without your involvement — it plans its own next steps, carries them out and launches the following ones, sometimes handing tasks off to further agents.
It sounds appealing and in some cases it genuinely works. But stay alert to one detail: when the model makers themselves promote that narrative, they have an obvious interest in it — loops running nonstop consume far more tokens, and tokens are their revenue. From my perspective, most everyday knowledge work doesn't need agents running nonstop at all. An open loop with a vague instruction is often just a way to burn through your session limit before you're back at the desk — with no guarantee the result will be better than after three well-thought-out instructions. Longer autonomy is a great tool for tasks that really are long and well defined. It isn't an argument for letting everything run on its own.
The takeaway rule
This launch nicely illustrates a rule I hold to: a model's capabilities are neutral — value or harm is created only by the judgment of the person who runs them. Anthropic wrote that principle straight into the architecture of its offering, splitting Fable 5 and Mythos 5 precisely along the line of safeguards. For you, the same rule works on a smaller scale: it isn't access to a more powerful model that will improve your work, but what tasks you put it to and with what head on your shoulders.
The next step is simple: before the window closes on 22 June, try Fable 5 on your own everyday tasks and note where the difference is real and where it's cosmetic. Then the decision about credits after 23 June will all but make itself.