Muse Wonder vs Claude Fable 5: Two Index Points for 40% of the Rate, and a Successor Complicating Both

Claude Fable 5 is A​nthropic's first Fable-class model, shipped June 9, 2026 at $10.00 per million input tokens and $50.00 per million output tokens. Muse Wonder is Meta Superintelligence Labs' frontier release of October 12, 2026, at $4.00 and $20.00. On Artificial Analysis' current Intelligence Index the older, dearer model reads 50 and the newer, cheaper one reads 52. But the reason this matchup is genuinely strange is a third model: Claude Fable 5.1, which replaced Claude Fable 5 on September 1, 2026 at the exact same $10/$50 list, scores 53, and cut the cache read from $1.00 per million tokens to $0.25. The model on the left of this comparison has already been superseded by something cheaper to run at the same sticker price — and its successor is not part of the matchup you asked for.

So this page does two things. It measures the matchup you asked for — Muse Wonder against Claude Fable 5 — and it tells you when that matchup is the wrong question, which here is most of the time.

The two models, and the one that stands behind each

Claude Fable 5 is A​nthropic's Mythos-class flagship, a capability tier the company places above its Opus class. It takes text, image and file input and returns text, serves a 1,000,000-token context window with a 128,000-token ceiling on a single response, and supports adaptive reasoning and structured outputs. It is still served, still priced at $10.00 and $50.00 per million tokens, and still carries a $1.00 per million cache read. Artificial Analysis flags it as superseded and points readers to Claude Fable 5.1.

Muse Wonder is Meta Superintelligence Labs' frontier release of October 12, 2026, announced as part of a two-checkpoint series. Only muse-wonder-1.0 has a per-token endpoint — on the Meta Model API and inside Muse Code — and it is the one every figure here describes. It carries a 1,048,576-token context window, a 943,718-token maximum output, an input surface covering text, image, speech and video, and mandatory reasoning with an effort ladder from minimal to max that defaults to medium. M​eta's second checkpoint, Muse Wonder Full, scores higher and is sold only through a Muse Code subscription, so it has no rate card to compare against.

Read those two paragraphs side by side and the shape of the comparison is already visible. One is a five-month-old model with a settled independent record; the other is a two-day-old release with a single evaluator's row. The price gap is 2.5× in M​eta's favour. The capability gap, on the only board that has measured both, is two points the other way.

The measured gap, and what it is built from

Artificial Analysis is the only evaluator with published rows for both models, and both were run at maximum effort on the same revision of the same index. That makes these directly comparable to each other in a way most vendor-versus-vendor numbers are not.

• Intelligence Index — Muse Wonder 52, ninth of 216 models; Claude Fable 5 50, twentieth of 216

• Cost per index task — Muse Wonder about $4.52; Claude Fable 5 about $8.75

• Output tokens consumed on the index run — both about 130 million, against a board median near 85 million

• List rate — Muse Wonder $4.00 in / $0.20 cached / $20.00 out per million; Claude Fable 5 $10.00 in / $1.00 cached / $50.00 out

• Context window and output ceiling — Muse Wonder 1,048,576 in and 943,718 out; Claude Fable 5 1,000,000 in and 128,000 out

• Independent serving measurements — Muse Wonder about 115 output tokens per second with a time to first token near 20 seconds; Claude Fable 5 about 66 tokens per second with a time to first token near 129 seconds

Two things in that list are worth pulling out. The first is the verbosity row: both models burn roughly the same number of output tokens to complete the evaluation, so the five-point list-price gap on output is not softened by one of them talking less. The cost-per-task ratio is close to the rate-card ratio, which is what you get when the two models write at similar length. The second is the latency row, where the two are not close at all — a two-second difference in time to first token changes what a workload feels like and what it can be used for interactively.

Five months of model generations are worth two index points

The most honest observation about this pairing is how small the capability gap is across a five-month release gap, and how much of the pricing gap comes from a decision rather than from the calendar.

M​eta priced Muse Wonder at 40% of A​nthropic's Fable rate. A​nthropic, for its part, did not cut the Fable price when it shipped Fable 5.1 — it lowered the cache read instead. Those are two different bets about what buyers pay for. M​eta's bet is that the headline rate is the number that shows up in a procurement conversation. A​nthropic's is that the cache read is the number that shows up on the invoice, because agentic workloads re-send large prompt prefixes constantly and the cached line is where the money actually goes.

For a workload with a stable prefix — a repository in context, a fixed system prompt, a tool schema that does not change — a $1.00 cache read against a $0.20 cache read is a 5× difference on the largest token category in the request. That number is larger than the 2.5× on the sticker, and it is the reason Claude Fable 5's position against Muse Wonder is worse than its list price suggests, while its successor's position is better.

Where the comparison changes with the successor in it

Claude Fable 5.1 scores 53 on the same index against Claude Fable 5's 50, and it lists at the same $10.00 and $50.00. It carries a $0.25 cache read rather than $1.00, and it supports a batch lane at $5.00 and $25.00 for asynchronous work. In other words, the successor is strictly better than the predecessor on every axis a buyer cares about and costs exactly the same per token.

That does not make the original uninteresting. It makes it useful in one specific way: it prices what a version bump is actually worth, in public. Three index points, a 4× cache read, and a batch lane — that is the entire delta between a flagship and its own refresh, and it is a reminder that on a frontier model the version number matters more than the generation label.

• Against Claude Fable 5, Muse Wonder leads by two index points and costs about 48% as much per completed index task

• Against Claude Fable 5.1, Muse Wonder trails by one index point and costs about 41% as much per index task, and its cache-read advantage narrows from 5× to 1.25×

If you are choosing a model to write into a production path today, the second comparison is the real one. The first is a useful benchmark of how much the market moved between June and October of 2026, and very little else.

Reasoning control, which is not a small difference for agents

The two models expose effort in different ways, and for anything with a cost budget the difference is operational rather than academic.

• Claude Fable 5 — adaptive reasoning with a documented fallback configuration, which means the model can decide not to spend effort on a request that does not need it, plus a batch mode for work that can wait

• Muse Wonder — mandatory reasoning. The ladder runs minimal through max and defaults to medium; there is no configuration that turns thinking off for a cheap classification call

That asymmetry cuts both ways. Mandatory reasoning is why Muse Wonder posts a strong agentic profile and a high output-token count; it is also why a high-volume pipeline of short, easy calls will spend more on it than the rate card suggests. If your traffic is uniform and hard, the default is a feature. If your traffic is a long tail where most calls are trivial, a model that can decline to think is worth more than two index points.

On the routing side this is the cleanest kind of split, because both models are callable through one key. OrcaRouter carries Claude Fable 5 as anthropic/claude-fable-5 at A​nthropic's own list price with 0% markup — provider list price passed straight through, so a rate change at A​nthropic is live here the same day — and more than 200 models sit behind the same credential. Muse Wonder is absent from our catalogue: a router can only route a model somebody is serving, and today the only serving path for M​eta's new checkpoint is M​eta's own API and Muse Code. A routing DSL that composes several models into one call is what makes a split like this practical rather than aspirational — cheap model for the easy calls, expensive one for the hard ones, automatic failover when an endpoint is unhealthy — and what it cannot do is conjure a second provider for a model that has one.

Which one to call, and which one to call instead

The rule that survives both readings of this page is short. Choose Muse Wonder when the workload is hard, long and agentic, when the prompt is too big to fit comfortably under a million tokens with room to spare, and when a 943,718-token maximum response is something your pipeline actually needs. Its rate card is the cheapest frontier list price published by any of the three major labs, and its cache read is the best of the three.

Choose against it when your traffic is dominated by short calls with little cacheable prefix, because mandatory reasoning is the tax you pay there and the sticker price hides it. And if the reason you opened this page was to decide between M​eta's new model and A​nthropic's Fable line, the honest answer is that you should be pricing Muse Wonder against Claude Fable 5.1, not Claude Fable 5 — the predecessor has been superseded in place, at no extra cost, and comparing against it flatters the wrong column.

The forward-looking version: everything above is a snapshot taken on October 13, 2026, when Muse Wonder was two days old and carrying one independent row. Whether the two-point lead over Claude Fable 5 survives a second evaluator, and whether M​eta ever publishes a per-token endpoint for Muse Wonder Full, are the two developments that would make this page worth rewriting rather than re-reading.