Muse Wonder vs Muse Spark 1.2: Meta Released a Flagship That Costs 4.7× the Model Below It

Muse Wonder costs $4.00 per million input tokens and $20.00 per million output tokens. Muse Spark 1.2 costs $1.25 and $4.25. Same lab, same 1,048,576-token context window, same text-out shape, and a price ratio of 3.2× on input and 4.7× on output — for a model released on October 12, 2026 that does not clearly outscore the August checkpoint it is priced above. On Artificial Analysis' current Intelligence Index, Muse Wonder (max) reads 52. Published third-party readings of Muse Spark 1.2 have landed between 54 and 57 depending on the revision and the effort setting used. That is not a comparison where the flagship wins and you pay for it; it is a comparison where you have to decide what you are buying the extra 3.7× for.

The complication is that Muse Spark 1.2 is no longer Meta's current Spark. Muse Spark 1.3 shipped on September 2 and Meta's own documentation now calls 1.2 "the previous version." So this page is really a three-way question wearing a two-name title: the cheap previous checkpoint, the current cheap checkpoint, and the expensive new flagship — all from Meta Superintelligence Labs, all reachable in different ways, all with the same context window.

Three months, three releases, one price band that did not move

Meta's Muse line has moved fast, and the rate card has not moved with it on the Spark side.

• Muse Spark 1.1 — the original, listed in our catalogue at $1.25 per million input tokens and $4.25 per million output, with text, image, video, audio and PDF input.

• Muse Spark 1.2 — released August 5, 2026, the same $1.25 and $4.25 on the Standard tier, Meta describing it as an updated checkpoint rather than a new tier. A second SKU, Muse Spark 1.2 Contributor, runs the same checkpoint at $0.10 in and $0.20 out on Meta's Contributor tier.

• Muse Spark 1.3 — shipped September 2, 2026, still $1.25 and $4.25, and now the version Meta recommends for new work.

• Muse Wonder — released October 12, 2026, at $4.00 in and $20.00 out, with a 943,718-token output ceiling and a second, higher-scoring checkpoint called Muse Wonder Full that is Muse Code subscription only and has no per-token price at all.

Read that as a set of decisions rather than a timeline. In three months Meta shipped four Spark-or-Wonder releases, moved the Spark price not at all, and then introduced a new name at triple the price of the model underneath it. The naming is the tell: Muse Wonder is not Muse Spark 1.4. It is a separate series, and Meta priced it as one.

The rate card, and what the reasoning setting does to it

• Input — Muse Wonder $4.00 per million against Muse Spark 1.2 at $1.25. 3.2×.

• Output — $20.00 against $4.25. 4.7×, and the line that dominates the bill in any agentic loop.

• Cached input — Muse Wonder $0.20 per million; Muse Spark 1.2 $0.15. Half a cent apart in absolute terms, which is worth noting because both vendors present caching as the discount that rescues the expensive model. On this pair it does not: the cache read is not where the 4.7× lives.

• Context — 1,048,576 tokens on both. Identical, and identical to the previous two Spark checkpoints as well.

• Output ceiling — Muse Wonder 943,718 tokens in a single response; Muse Spark 1.2's published output ceiling is not stated on our catalogue row at all, and Meta's documentation for the Spark line does not put a comparable figure on it. Treat that as unknown rather than equal.

• Reasoning — Muse Wonder's is mandatory, with an effort ladder from minimal to max defaulting to medium. Muse Spark 1.2 supports configurable reasoning effort, and Meta's own description of the family is that effort is a dial rather than a floor.

• Modalities — text, image, video, audio and PDF in on Muse Spark 1.2; text, image, speech, video and file in on Muse Wonder. On input surface these are the same model class, which is the point: nothing here explains a 4.7× output price.

What the independent board actually says, and why the numbers disagree

This is the part to read slowly, because the published figures for Muse Spark 1.2 do not agree with each other and the reason matters.

Artificial Analysis' analysis at Muse Spark 1.2's launch scored it 54 on the Intelligence Index in its xhigh configuration, an increase over the previous checkpoint. Other readings of the same checkpoint have been published at 57, and our own notes on the family carry both figures. The discrepancy is not a mystery and it is not a vendor-versus-independent fight — it is the index itself moving. Artificial Analysis re-scored its board during September 2026, replacing easier component benchmarks with harder ones and raising the weight on private held-out tasks, and scores fell across the frontier as a result. A number measured before that re-scoring and a number measured after it describe the same unchanged model on two different rulers.

So the defensible statement is a band rather than a point: third-party readings of Muse Spark 1.2 have been published between 54 and 57, and the comparison that matters is against Muse Wonder's 52 measured on the current revision. Which means the honest summary of this matchup is that on independent measurement the August checkpoint is at least level with, and possibly ahead of, the October flagship — while costing a fifth as much on output.

Meta's own vendor-reported benchmark table points the other way, which is exactly what a vendor table about its own new flagship is expected to do. Muse Wonder is listed ahead of Muse Spark 1.2 on the long-horizon professional suites, and its Full checkpoint dominates those rows outright. Every figure is from Meta's harness and none of it has been independently reproduced. It belongs on the list of reasons to run your own evaluation, and it is not evidence that the current board is wrong.

The case for the cheaper model being the default

There is a version of this comparison that Meta's pricing invites and that most teams should probably take seriously: for a large class of agentic work, Muse Spark 1.2 or 1.3 at $1.25 and $4.25 is the sensible default and Muse Wonder is the escalation.

The argument is not that the cheaper model is better. It is that the two are close enough on measured capability — within a few points on the same board, in a direction that does not clearly favour the expensive one — that the price difference should be spent on where the workload actually hurts. A coding agent that spends most of its tokens on routine edits and a small fraction on the hard reasoning step gains more from running 4.7× the volume at the cheap rate than from running the same volume at the flagship rate. Meta's own Contributor tier pushes that further: the same Muse Spark 1.2 checkpoint at $0.10 in and $0.20 out is roughly 40× cheaper than Muse Wonder on input, which is a difference that changes what is economically possible rather than merely what is affordable.

What Muse Wonder has that the Spark line does not is narrower and more concrete than a benchmark lead: the 943,718-token single-response output budget, and a Frontier-lab flagship positioning that comes with a subscription-gated higher tier. If your work generates long artifacts in one turn, that output ceiling is the entire argument. If it does not, the argument is much thinner.

One more distinction to keep straight, because the naming makes it easy to fumble: Muse Wonder Full — the higher-scoring half of the October release — is not sold per token at all. It is described by Meta as Muse Code subscription only, on plans published at $50 a month for Power Usage and $200 for Maximum Usage. Nothing on this page's rate-card comparison applies to it, and a team that reads a vendor table showing Muse Wonder Full ahead of everything else and then looks at the $4.00 rate is looking at two different models.

Cascading inside one family, which is the real answer here

The honest conclusion of a same-lab comparison this close is that the choice is not binary, and the mechanics of running both sides make that practical. Muse Spark 1.2 is on OrcaRouter as meta/muse-spark-1.2 at Meta's Standard-tier rates — $1.25 input, $4.25 output, $0.15 cache reads, the 1M context window and the full multimodal input surface passed through — and Muse Spark 1.1 is there at the same price with identical modalities. That makes the cheap half of this comparison reachable from one key today.

Muse Wonder is not on OrcaRouter. There is no meta/muse-wonder entry in our catalogue and we will not claim one; Meta's own API and Muse Code are where it runs. What that leaves is a routing problem with a clean shape: put the cheap Spark checkpoint on the default path and escalate only the requests that need the flagship's output budget. OrcaRouter's routing DSL composes several models into a single call, so a cascade can be expressed as configuration rather than as application logic — and automatic failover means a provider incident degrades a request to the next candidate instead of failing it. Across 200-plus models on one OpenAI-compatible endpoint with provider list prices passed through untouched, the cost of testing whether the escalation is worth it is a model-name change, and the day a provider we route starts serving Muse Wonder it becomes one of the candidates in that cascade rather than a separate integration.

Worth saying plainly as well: this page is not an argument against the flagship. It is an argument against paying 4.7× for output without measuring what the 4.7× returns on your own traffic — and within one lab, where the input surface and the context window are identical, that measurement is unusually easy to set up.

The short version

If you generate long outputs in single turns, or you need speech and video going into the model without a transcription step ahead of it and a 943,718-token answer budget to go with them, Muse Wonder is the model in this family and the 4.7× output premium is the price of a capability the Spark line does not advertise. If your agentic traffic is mostly routine work with occasional hard steps, Muse Spark 1.2 or 1.3 at $1.25 and $4.25 will absorb several times the volume for the same money, and the independent board does not give the flagship enough of a lead to justify moving. Build the cascade, put the cheap checkpoint first, and let your own token counts tell you how often the escalation fires — that number, not this page's, is the one that decides it.