Fable 5.1 Is In Your Max Plan. It Is Not Free.

If you pay $200 a month for Claude Max 20x, Fable has been part of your subscription since 20 July 2026. Up to half of your weekly usage limits may go to Fable at no extra charge. Read that precisely: it is a ceiling inside the allowance you already have, not a second pool bolted on top. A lot of people who wrote Fable off as the expensive model never noticed it moved at all.

Fable 5.1 and Mythos 5.1 shipped today, 1 September 2026, nearly three months after Fable 5. The sticker price did not move: $10 per million input tokens, $50 per million output — exactly twice Opus 5 on both. What did move is cache reads, cut 75% from $1.00 to $0.25 per million, which is half what Opus 5 charges for the same thing. So Fable is twice the price on fresh tokens and half the price on cached ones.

So the interesting question is not whether Fable is expensive. It is whether the price you actually pay, in dollars or in weekly allowance, is anywhere near double.

What the 50% cap actually means

Anthropic's support documentation now covers Fable 5 and Fable 5.1 explicitly. On Max and Team Premium you may spend up to 50% of your weekly usage limits on Fable models at no extra cost. The same document is blunt about which pool that is: Fable models draw from your plan’s regular weekly usage limits and use them faster than other Claude models, and you can never use more than your weekly limit. There is one bucket. Exhaust it and switching to Fable rescues nothing, because Fable was drinking from the same bucket the whole time. Hit that ceiling and you switch model or wait for the weekly reset. Continuing on Fable is billed to usage credits at API rates, and only if you have enabled them.

Pro received a one-time $100 credit when Fable 5 moved to this arrangement in July. Fable 5.1 came with no equivalent promotion for anyone. If you are on Pro, Fable is pay-as-you-go at full API rates and this article's answer for you is short: use Opus.

One practical detail that will waste an afternoon if you miss it. Fable 5 needs Claude Code 2.1.170 or later. Fable 5.1 needs 2.1.250 or later. The model ID is claude-fable-5-1.

Note what the cap is denominated in. Not dollars — allowance. Anthropic does not publish absolute token budgets, so the useful figure is a relative one: Fable weighs roughly double an Opus session against that allowance. The API rates and the subscription burn point the same direction. A Max subscriber pays the 2x twice over, once in each currency, and only escapes it if Fable genuinely needs fewer sessions to finish the same work. That is why the honest unit of measurement here is allowance consumed per finished task, not tokens and not dollars.

The arithmetic, out loud

Fable 5.1 also charges $12.50 per million for a five-minute cache write and $20.00 for the one-hour variant. Two things fall straight out of this table. Fable 5.1 standard matches Opus 5 fast mode on input and output, so if you already run Opus in fast mode you are paying Fable rates today — though Fable still reads cache at half the price. And at equal volumes Opus 5 batch is half the price of Fable 5.1 batch, so delay-tolerant work stays cheaper on Opus.

Anthropic's own framing of the cache cut is that it reduces Fable's effective cost by roughly 25% on typical workloads and up to about 45% on highly agentic ones. Read that carefully — those are savings against Fable 5, not against Opus. The company is not claiming Fable is cheaper than Opus, and neither am I.

A worked bill for one long agent session

Assume a session that burns 2M ordinary input tokens, 400K output tokens and 40M cache reads — an agent repeatedly rereading a large cached repository. These are illustrative numbers, not measurements, and cache-creation charges are excluded.

On this illustrative mix, at identical token volumes, Fable costs 25% more rather than 100% more. The cache line does all the work: it is the only row where Fable is cheaper, and in agentic coding it is usually the biggest row. This proves one traffic mix, not a universal premium, and it excludes cache-creation charges.

The break-even is worth stating as algebra, because it lets you compute your own. With I million input, O million output and C million cache reads at equal volumes, the two bills are equal when C = 20I + 100O. For this example that is 80M cache reads. We assumed 40M, so Fable stays more expensive on volume alone — until token efficiency enters.

And that threshold is low. If input, output and cache reads all fall in the same proportion, Fable reaches parity at 80% of Opus's volume. Halve all three and the bill lands at $25, 37.5% under Opus. Real agents will not reduce every category proportionally, so treat 80% as the shape of the answer rather than the number.

The claim everything hinges on, and the people who dispute it

Dan Shipper at Every reports Fable 5.1 running roughly twice as fast as Opus 5 while using about half the tokens. If that reproduces on your work at comparable acceptance quality and retry rates, the modelled bill favours Fable: double the per-token price times half the tokens is a wash, and the cheaper cache reads then push you below Opus.

Treat it as a partner report, not an independent measurement.

Because on Hacker News, developers using it are saying something close to the opposite. The recurring complaint about Fable is verbosity, not efficiency. One commenter reported cancelling their Max subscription outright, saying Codex was more succinct. Another summed the economics up as Fable models taking more tokens while living under a 50% limit or running on credits. On 5.1 specifically, one reader described the prose as denser than Fable 5's, with longer sentences and fewer breaks. That is a style observation, not a token measurement. Nobody in that thread reported the doubled speed. These are anecdotes from a self-selected group, not a representative sample.

Both things can be true. A model can produce chattier prose while completing an agentic task in fewer total turns, and total turns is what actually bills. But you should not spend money or allowance on the assumption that it does.

The benchmarks are more mixed than the headlines

"Beats Opus 5 on most benchmarks" is accurate and slightly misleading. On SWE-bench Verified — the one most people still anchor on for coding — Opus 5 is still ahead, 96.0% to 95.5%. Fable's wins are concentrated in agentic and terminal-driven work, which is a real and relevant category, just not the one the headline implies. Its Terminal-Bench-Science score more than doubled from Fable 5's 24.7%, which is the single biggest jump in the release. One caveat on the Terminal-Bench 4.0 row: sources disagree on Opus 5's figure, with at least one reporting 52.3% rather than 42.0%, so treat that gap as directional.

Headless and interactive share one pool

This one catches people. On 14 May 2026 Anthropic announced that Claude Agent SDK and claude -p headless usage would leave the subscription pools on 15 June and move to a separate monthly credit balance at API rates — roughly $20 for Pro, $100 for Max 5x, $200 for Max 20x, no rollover.

That change was cancelled before it took effect. As of today, interactive terminal sessions, claude -p, Agent SDK runs and third-party app usage all draw from the same subscription allowance. There is no separate agent credit pool.

For anyone running unattended agents: a night of headless runs consumes the same allowance you need for the next morning's interactive work, and Fable's 50% cap applies to that shared pool rather than getting its own headless budget. If you sized your automation around the announced credit system, resize it. Related: Your Agents Eat Your Subscription.

Run the one-week test instead of guessing

Turn Fable 5.1 on and keep standard Opus 5 as the control — not fast mode, which shares Fable's pricing and answers a different question. Alternate comparable tasks between them for a week rather than sending all the hard work to one model. Track this:

Mythos 5.1, briefly, because you cannot have it

Mythos 5.1 is the same underlying model as Fable 5.1 with the safety classifiers lifted. Access runs through Project Glasswing and is restricted to vetted organisations under two programmes: the Cyber Verification Program for defensive security work, and the Life Sciences Verification Program. It is not available to ordinary individual accounts, and it has no bearing on the Max decision. It is worth noticing only as a signal about where Anthropic thinks a classifier-free model belongs, which is inside a small number of verified institutions.

Bottom line

If you are on Max 20x, the answer that was true six weeks ago is no longer true. Fable is not something you buy separately — it draws on the limits you already pay for, up to half of them. And on a cache-heavy mix like the one above, the premium over Opus works out around 25% rather than the 100% the sticker implies.

That does not make it the default. The claim that would make it the default — half the tokens, twice the speed — comes from an Anthropic partner, while some developers posting publicly report the opposite about verbosity. No independent controlled comparison has been published.

So use the included share. Run it for a week against standard Opus on real tasks and measure allowance consumed per accepted result, not raw tokens. On the worked example Fable needs roughly a 20% proportional token reduction to break even. If it delivers the claimed 50%, make it your interactive default until the cap and let Opus take the rest. If it delivers nothing, you have spent shared allowance but no extra dollars — provided you disable paid extra usage first, so you cannot roll past the cap onto credits without noticing.

The 50% is a ceiling, not a gift. Make Fable earn the share of the allowance it takes.