This morning Anthropic released Claude Fable 5, the first publicly available model from its new Mythos tier. The Mythos class sits a step above the Opus models that until today were the top of the range. It's priced at $10 per million input tokens and $50 per million output tokens, which makes it the most expensive mainstream model on the market, roughly double Opus 4.8. Two details matter more than the benchmarks. First, Anthropic says the model's lead grows as tasks get longer and more complex. That tells you exactly what it's for and what it isn't. Second, it's included free on paid Claude plans until 22 June. That's a fortnight to find out whether the dearest model in the world is worth it on your work, using your documents, at zero marginal cost. Here's the test I'm running this week, and the trap I'd avoid.
New model launches have become a bit like phone launches: a day of breathless coverage, a benchmark chart, and then everyone goes back to work. I nearly filed this one in that drawer. Then I read the pricing page and the small print, and changed my mind, because this launch contains a rare thing — a free, time-boxed experiment that any business on a paid Claude plan can run with no setup at all.
The facts first. Claude Fable 5 launched today, 9 June. It's a “Mythos-class” model — Anthropic's new tier above Opus — and the company is unusually plain about the trade it's making: the full-strength Mythos 5 is restricted to vetted organisations, while Fable 5 is the same underlying model wrapped in extra safeguards for the rest of us. Ask it something in a handful of high-risk areas — cybersecurity exploits, biology, chemistry — and it quietly hands the question down to Opus 4.8 instead.
And the price: $10 in, $50 out, per million tokens. For comparison, Opus 4.8 — the model that was Anthropic's flagship until this morning — sits at $5 and $25. You are being asked to pay double for the new top tier.
What doubling the price buys you
The line in Anthropic's launch material that I'd underline is this: the longer and more complex the task, the larger Fable 5's lead over its other models. That's not marketing filler. It's a precise description of where the money goes, and it maps onto something I see in client work every month.
Most of what a small business asks an AI to do is short-haul. Draft this email. Summarise this call. Categorise these expenses. On short-haul work, the difference between a very good model and the best model in the world is somewhere between small and invisible. Paying double for invisible is how AI budgets quietly bloat. I wrote in April about auditing my own stack and cutting £125 a month of exactly this kind of spend, and nothing about today changes that discipline.
But some work is long-haul: the multi-step jobs where the model has to hold a plan in its head across an hour of grinding. Reconciling a year of messy supplier data. Reviewing a 200-page contract pack against your standard terms. On those jobs, a model that's 10% better per step compounds into something dramatically better by the end. Each small error it doesn't make is an error nobody has to find later.
The fortnight that makes this worth your attention
Here's the detail that turns a launch announcement into something actionable: Fable 5 is included on Claude Pro, Max, Team and Enterprise plans at no extra cost until 22 June. After that, using it draws on usage credits. Anthropic is doing this to get adoption data, obviously. But for you it's a fortnight in which the most capable generally available model on the planet costs the same as the plan you're already paying for.
I'd treat it like a free trial of a very expensive employee, and structure it accordingly. This week I'm running the same exercise myself and with two clients:
- Pick your three ugliest recurring tasks. Skip the easy wins. Pick the jobs current AI gets 80% right and you've stopped delegating because fixing the 20% eats the saving. For my own ops, that's a monthly multi-source reporting job that still needs an hour of my checking. For a jewellery e-commerce client, it's matching supplier invoices to a catalogue that three people maintain in three different styles.
- Run them side by side. Same prompt, same documents, current model versus Fable 5. Don't eyeball it. Count the corrections you have to make on each output. Corrections are the honest currency here, because your time fixing AI output is the real cost that never shows on the invoice.
- Do the maths before 22 June. If the correction count drops enough to matter, work out what the task would cost at $10/$50 and compare it to the hours saved. If it doesn't drop, you've just learned — for free — that your stack is already good enough, which is worth knowing every time a launch tries to convince you otherwise.
The safety routing is a feature, not an insult
Some of the early commentary has treated the high-risk fallback — Fable 5 declining certain topics and handing them to Opus 4.8 — as a limitation to grumble about. For a typical business user I'd argue the opposite. Anthropic shipped this model days after publicly warning that frontier AI is becoming dangerous in specific, technical ways. A vendor that gates its strongest model's riskiest capabilities, and is open about doing so, is behaving exactly the way you'd want every vendor on your stack to behave. If your business needs were ever going to collide with those gates, you'd know already, and you'd be in the vetted tier.
One caveat before you re-platform anything
June is shaping up to be the busiest model-release month of the year. Gemini 3.5 Pro is confirmed as imminent, and there's credible evidence a new Claude Sonnet lands within weeks. It'll likely be cheaper than Fable 5 and closer to it in capability than today's mid-tier. So the worst response to this launch is to tear up your workflows and rebuild them around the new top model on day two. Prices and rankings in this market have a shelf life measured in weeks; decisions about your operations should not.
Run the experiment, keep the receipts, and decide on 22 June with your own numbers in front of you. The model is the most expensive ever sold. The fortnight to find out whether that matters to your business costs nothing. A free answer to a £several-hundred-a-month question is the best deal in AI this week.