Claude Fable 5 Pricing 2026:
What It Costs and Is It Worth It?

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens on the API, double what Claude Opus 5 and Opus 4.8 cost. On a subscription, what you pay depends entirely on which plan you are on, and that answer changed on 20 July 2026.

Two questions come up more than any other: how much does Fable 5 cost, and is Fable 5 worth it. This article answers both, in that order. (You will see the model written both ways, Claude Fable and Claude Fable 5. There is only one, and Claude Fable pricing means the figures below.)

Fable 5 has had the most chaotic pricing history of any model Anthropic has shipped. Free for three days, switched off by a government directive for eighteen, redeployed, then given three separate deadline extensions before Anthropic finally settled it. This article is the settled version: what it costs on every plan, what UK buyers actually pay, why it eats your usage limits twice as fast as Opus, and whether it beats the model you already have.

What Claude Fable 5 costs

There are two separate prices, and most of the confusion online comes from people quoting one when they mean the other.

On the API, Fable 5 is $10 per million input tokens and $50 per million output tokens. That is a flat, published rate with no plan attached, and it is exactly double the $5/$25 that both Claude Opus 5 and Claude Opus 4.8 cost.

If none of that means anything to you, here is the translation. A token is a chunk of text a little shorter than a word, so a million tokens is roughly 750,000 words, or about eight full-length novels. "Input" is everything you send the model, including the files and conversation history it reads. "Output" is what it writes back, and it costs five times as much, which is why the second number matters more than the first on any job that generates a lot of code. In practice: a substantial afternoon's coding session might run through a few million tokens, so you are talking tens of dollars, not cents, and not hundreds.

On a subscription, you are not buying tokens, you are buying weekly usage limits, and Fable 5 draws against those limits rather than billing you separately. Whether it draws against them at all depends on your plan.

If you are on the API rather than a subscription, the only number that matters is what a real call actually costs you, and you can print that from the response rather than guessing. Every response carries a usage object with the exact token counts:

import anthropic

client = anthropic.Anthropic()

resp = client.messages.create(
    model="claude-fable-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarise this changelog in three bullets."}],
)

# Fable 5 bills $10 per million input tokens and $50 per million output.
# Output is 5x the price of input, so it dominates on anything that writes code.
u = resp.usage
cost = (u.input_tokens / 1_000_000 * 10) + (u.output_tokens / 1_000_000 * 50)
print(f"{u.input_tokens} in, {u.output_tokens} out, ${cost:.4f}")

Swap claude-fable-5 for claude-opus-5 and halve both multipliers to compare the same job on the cheaper model. Running that on a handful of your own real prompts will tell you more than any benchmark table, including this one.

So why is Fable 5 so expensive? Two reasons, and only one of them is about capability. It is a bigger model that costs Anthropic more to serve, which is the honest half. The other half is scarcity: Anthropic priced and rationed Fable during a period when it plainly did not have the GPUs to serve it at Opus volumes, which is also why the Claude Fable 5 price moved four times in six weeks. Price on a frontier model is partly a queue management tool, not just a margin.

The sticker shock needs one qualifier. Per-token price is not per-task price. A model that solves a job in one attempt can cost less overall than a cheaper model that needs three runs and an hour of your corrections. Finout's cost analysis makes the same point from the other direction: for routine, well-defined work like summarising and extraction, the 2x premium is hard to justify. Both things are true at once. The premium pays for itself on hard jobs and wastes money on easy ones.

Claude Fable 5 pricing on each plan

This is the table that actually answers the question, and it has been correct only since 20 July 2026. What a Claude Fable 5 subscription gets you depends entirely on which row you are in.

Anthropic's own support documentation is the authority here, and it is worth reading if you are on Team or Enterprise, where the split follows seat type rather than plan name: Premium seats get the 50% inclusion, Standard seats get credits.

How we got here is a useful warning about trusting any pricing article on this subject, including the first version of this one. The original plan was thirteen free days ending 22 June. The suspension killed that. The replacement deadline was 8 July, extended to 12 July after complaints, extended again to 19 July, and then abandoned entirely in favour of permanent Max inclusion. Four deadlines in six weeks.

Simon Willison, who had been tracking the saga closely, put the commercial logic plainly when the permanent change landed.

That question answers itself, which is presumably why Anthropic blinked. Willison attributes the reversal to competitive pressure from GPT-5.6 Sol and Kimi 3, and notes the original credits-only plan was driven by compute capacity rather than greed. Both readings fit the evidence. A company genuinely short of GPUs and a company protecting its highest-margin tier look identical from outside.

Claude Fable 5 UK price and availability

Fable 5 is available in the UK, on identical terms to the US, and has been since Anthropic redeployed it globally on 1 July 2026 after the export-control directive was withdrawn. There is no separate UK rollout to wait for and no regional model picker difference.

Pricing is where UK buyers get caught out. Anthropic has no UK price list. Unlike Google and Microsoft, which publish VAT-inclusive sterling prices, Anthropic bills in US dollars and adds 20% UK VAT at checkout. What lands on your card is the dollar figure, plus VAT, converted at your card provider's rate on the day. Two people on the same plan can pay different amounts in the same month.

At roughly $1.33 to the pound, the sterling cost works out at about £18 a month for Pro, £90 for Max 5x and £180 for Max 20x, VAT included. We pay the Max 20x figure ourselves, and £180 is the number that shows up on our statement, which is a reasonable sanity check on the arithmetic. Watch the exchange rate rather than the headline price: a five-cent move in sterling is worth about £7 a month on Max 20x.

If you are weighing a Claude subscription against the alternatives as a general-purpose tool rather than a coding one, our comparison of the best AI subscriptions covers all the major UK options with the same VAT treatment applied consistently.

Why Fable 5 draws down usage faster

Open the model picker in Claude Code and you get a warning along the lines of: Fable 5 is the most capable model and draws down usage much faster than Opus 4.8. Anthropic's current wording puts a number on it, "draws down usage about 2x faster than Opus". Judging by how many people paste that sentence straight into Google, the warning raises more questions than it answers. Here is what it actually means.

The 2x is not a rough estimate or a penalty. It falls directly out of the token price. Subscription limits are metered against what the tokens would have cost on the API, and Fable 5 costs exactly double Opus per token in both directions. So a task that would consume a quarter of your weekly allowance on Opus 5 will consume half of it on Fable 5, for identical work and identical output.

Now combine that with the 50% cap on Max plans, because this is the part nobody spells out. You may spend up to half your weekly allowance on Fable 5. Fable spends it twice as fast. So the Fable half of your week buys about a quarter of a week's worth of Opus-equivalent work. The 50% cap sounds generous and behaves like 25%.

That maths is worth doing before you switch your default model, because the failure mode is unpleasant. Run Fable as your everyday driver on Max and you will hit the Fable ceiling around midweek, drop back to Opus with a chunk of your general allowance already spent, and find your limits tighter than a normal week for reasons that are not obvious at the time.

Two practical consequences. If you use Fable, use it deliberately for specific jobs rather than leaving it selected. And if you are on Pro, where Fable runs on credits rather than limits, the drawdown question does not apply to you at all: you are simply paying API rates, and the meter is a lot easier to read.

What Claude Fable 5 actually is

In April we wrote about Claude Mythos, the model Anthropic built and then refused to release because it could find and chain zero-day exploits on its own. We predicted the reasoning gains would "filter into the product line" without the dangerous tooling. That is exactly what Fable 5 is.

Anthropic's announcement describes two models built from the same base. Claude Mythos 5 keeps the offensive cybersecurity capability and goes only to vetted partners through Project Glasswing. Claude Fable 5 is the public release: identical intelligence, with classifiers that detect restricted topics and route those requests to Opus 4.8 instead. It has a one-million-token context window, ships in Claude Code, the Claude apps, and the API as claude-fable-5, and it is not a speed upgrade. It is deliberately the opposite.

The fallback is worth understanding properly, because we got it wrong first time round. We originally described the routing as silent. It is not. Anthropic's documentation on model switching states that "you'll see a notice explaining that the model switched, and the response will be labeled with the model that answered". Four things trigger it: offensive cybersecurity, most biology and chemistry, attempts to extract the model's reasoning, and frontier LLM development work such as kernel design or distributed training.

One detail in that documentation deserves more attention than it gets. The safety checks review "everything the model reads, not just your latest message", including memory, connector content, web search results and files. So a fallback can be triggered by something in a document you attached rather than anything you typed. If you work anywhere near security or life sciences, that is worth knowing before you build a workflow on Fable 5.

What Claude Fable 5 is like to use

On 10 June, the day after launch, we used Fable 5 in Claude Code to build a new free tool for our sister site: the AI Bot Checker. The brief was outcome-shaped rather than a spec: an easy-to-understand checker that tests whether AI bots are blocked from a website, including robots.txt analysis. What it should do, and what using it should feel like. Nothing about how to build it.

Fable read through the existing codebase and the tools already on the site, then built the whole thing. Every major AI crawler tested two ways (robots.txt rules plus live requests sent with each bot's user agent), a bot-by-bot results table, an A to F grade, and a ready-to-send message for your hosting provider when something is blocked. Functionality and design landed right first time. No correction pass. In three years of building with these models, that had not happened to us on a job that size before.

The part that took the longest to specify, and that Fable got right without being told, was the crawler list itself. Deciding which bots to test is a judgement call rather than a lookup: some are indexers, some fetch a page live when a user asks a question about you, and two are not crawlers at all but robots.txt switches that control training. Here is the full set it checks, which doubles as a reasonable reference for anyone auditing their own robots.txt.

The two -Extended tokens are the ones that trip people up, and the tool marks them "not crawlable by name" for a reason. Google never crawls as Google-Extended. It crawls with Googlebot and honours the Google-Extended token in robots.txt as a separate permission for Gemini training and grounding. Apple works the same way with Applebot and Applebot-Extended. So there is no request you can send them and nothing a live test could prove: they are permission switches, not user agents. Blocking them does not remove you from search. We cover that distinction properly in our guide to whether your website is blocking AI.

The live test deserves one honest caveat, which the tool states on its own results page. Its requests identify as each bot but originate from our servers, not from OpenAI's or Anthropic's networks. So a rejection tells you your site or host is filtering traffic that claims to be that bot, which is usually what you want to know. But some security layers reject impostors while letting the verified bot through on a signed IP range. A failed live test is a strong signal to investigate, not proof that the real crawler is blocked, and any honest reading of the result has to hold that distinction.

One name is conspicuously absent. xAI publishes no crawler documentation, and Grok's observed fetches identify as an ordinary browser rather than a named bot. With no token to match in robots.txt and no user agent to test, there is nothing a checker or a firewall can reliably verify, so we leave it out rather than report a result we cannot stand behind.

Two things stood out in daily use. It is slower than Opus per response, sometimes noticeably. And it reads far more before it writes: more files opened, more existing patterns studied, before the first line of code appears. The slowness buys something. Fewer wrong turns, fewer "no, not like that" corrections, less of your time spent steering.

We were not the only ones seeing this pattern.

"The challenge is finding tasks that it can't do" is the line that made me put my coffee down, because it matched the day I had just had. I went looking for the catch in our one-shot build for a good hour: an edge case missed, a layout that breaks on mobile, anything. I didn't find one. Willison also buried a number in that post that matters more than any benchmark: he burned through $110.42 of Fable usage in a single day, inside his $100-a-month subscription. That was the free window doing its job, and a preview of what the meter looks like when it switches on.

Fable 5 vs Opus 5 vs Opus 4.8

When we first published this, the comparison that mattered was Fable 5 against Opus 4.8. Opus 5 landed on 24 July and made that framing obsolete.

Opus 5 wins Frontier-Bench by 9.6 points and GDPval-AA v2 by 114, matches Fable on CursorBench 3.2 to within half a point, and beats it on OSWorld 2.0 computer use at roughly a third of the cost. Fable leads exactly one row on this table, SWE-bench Pro, by about a point (80.3% against 79.2%). One point on one benchmark, at double the price and double the drawdown, is not a reason to change your default.

When Fable 5 is still the right call over Opus 5

I will be honest that this took me a while to work out for myself. Fable came back on our Max plan, I had no clear rule for when to reach for it, and because it eats the weekly allowance twice as fast I mostly just left it alone. If you are in the same position, here is the rule I settled on.

Ask whether you are going to be watching. That one question does most of the work, and it explains why the benchmark table looks self-contradictory at first glance.

Frontier-Bench is a one-shot test: hard problem in, answer out, scored. Opus 5 wins it because it checks its own work and questions the brief before acting, which is what its whole launch was built around. When you are in the loop reviewing each step, you and Opus 5 are both doing error correction, and paying double for a one-point ceiling is waste.

Long-horizon autonomous work is a different shape of problem. Anthropic's guidance is that Fable 5 remains its recommendation for agents that run for days, planning and self-correcting across very long trajectories with no human in the loop. On a run of thousands of steps nobody is there to catch a wrong turn at step 40, and a small per-step edge compounds into a very different outcome by step 4,000. That is where a raw capability ceiling stops being a rounding error and starts being the whole game.

So spend Fable allowance when:

• You are going to brief it and walk away. A migration, a large refactor, a build you want to come back to finished rather than supervise line by line.
• Being wrong is expensive to unpick. Not wrong in a way you spot immediately, but wrong in a way that propagates quietly through forty files before anyone notices.
• The job genuinely sits at the top of the difficulty range. That one-point margin is real, but it only shows up on the hardest problems. On ordinary work the two models are indistinguishable, and you are paying double for a difference you cannot perceive.

Stay on Opus 5 for everything else, which will be most things: conversational work, iterative work, anything you review as you go, and anything you could not confidently call hard.

The same job, briefed two ways. The difference is not prompt wording, it is the shape of the brief, and it falls straight out of that same question. The brief that produced the AI Bot Checker was, in substance: build an easy-to-understand checker that tests whether AI bots are blocked from a website, including robots.txt analysis, and here is what using it should feel like. Complete on the outcome. Silent on the method. It had to be, because I was not going to be sitting there to answer anything.

That last row is the single most useful change to make. "Ask me if anything is unclear" is dead weight on an unattended run: it either stalls the job or gets silently ignored. Swapping it for "decide, and tell me what you decided" gives you the same information at the end, in a list you can review, without the model waiting on a reply that is never coming.

The budget reality on Max. Half your weekly allowance, spent at double rate, is about a quarter of a week of Opus-equivalent work. That is two or three substantial jobs a month, not a daily driver. So if you cannot name the specific job before you switch models, that is your answer: don't. Leaving Fable unselected by default is the correct behaviour on a Max plan, not a failure to get your money's worth.

Not everyone reads Anthropic's tiering charitably, and the loudest sceptic runs the competition.

He said that in April, about the Mythos rollout that Fable 5 extends, and I will admit it lodged in my head while writing this. The cynical reading writes itself: invent a scary tier, gate it, then sell the safe version at double price. What softened my cynicism was watching Anthropic give the thing away on Max in the end, after four deadlines and a lot of public complaining. That looks less like a shelter sales pitch and more like a company short of compute, sampling demand, and then losing the argument with its own customers.

Is Claude Fable 5 worth it?

So, is Fable 5 worth it? The answer now depends on your plan, which is a better question than the one this article originally asked.

• On Max or Team Premium, Fable 5 costs you no extra money, so it is a question of allowance rather than budget. Apply the walk-away test: if you will be watching and reviewing, use Opus 5. If you are briefing a long job and leaving it to run, that is what the Fable allowance is for. Keep Opus 5 as the default for everything else.
• On Pro or Team Standard, Fable means buying usage credits at API rates. Spend the one-time $100 credit on your hardest backlog item to see what it does, then be honest about whether the result justified $10/$50 per million tokens. For most Pro subscribers it will not.
• If you are choosing purely on capability per pound, Opus 5 is the better buy. It wins on Frontier-Bench, matches Fable on CursorBench, costs half as much and draws down half as fast.

Our own position has moved since June. Opus 5 is now the daily default in Claude Code, and Fable comes out for the kind of job the AI Bot Checker was: big, well-briefed, self-contained, worth getting right first time. On current evidence that is perhaps two or three jobs a month, not twenty. The difference from June is that it now costs us nothing extra to make that call, which is the one genuinely good outcome from six weeks of pricing chaos.

One caveat before you commit real work to it. The safety fallback can fire on content you did not write, because the classifiers read your attachments, memory and search results too. You will be told when it happens. But if your work sits near security or life sciences, expect to meet Opus 4.8 more often than the 95% figure suggests.

What this means for your website

There is a thread connecting a frontier coding model to the work this site usually covers. The tool Fable built for us in one shot exists because AI crawler access has become something ordinary site owners need to check. Models this capable are increasingly the things reading your website: fetching pages, comparing claims, deciding whether your business is the one to recommend. We walked through that pipeline in how AI search actually works, and our server logs show AI crawlers reading machine-readable files daily.

The blocking problem is real, not theoretical. The first day we ran the new tool against live sites, one scored a D: ClaudeBot, Bytespider, and CCBot all rejected at server level with 403s, invisibly to the owner.

A model that can one-shot a working web tool can certainly parse your homepage. What it cannot do is invent facts you never published. AI Visibility Checking exists to verify the inputs are there: that crawlers can reach you, and that files like llms.txt and developer-ai.txt state who you are in a form these models can use. The quick-start guide covers the full set in an afternoon. As the models get sharper, ambiguity stops being neutral. It becomes the reason you are left out of the answer.

Frequently asked questions