Meta (Meta Superintelligence Labs)
Muse Wonder
Meta priced Muse Wonder at $4/$20 per million tokens, one flat rate on a 1,048,576-token window. Artificial Analysis's max-effort runs, checked 2026-10-13, put Humanity's Last Exam at 50.09 — a tie with Muse Spark 1.3's 48.7 inside the 2-point band — and Terminal-Bench 4.0 at 52.02 against Spark's 33.33 on the same runner, past that benchmark's 12.4-point band. Both of Muse Wonder's two scores are independent.
Muse Wonder benchmarks and pricing, every number sourced: 2 tracked Muse Wonder benchmark scores (2 independently run, 0 still resting on a vendor’s own claim), priced at $4.00 per million input tokens and $20.00 per million output.
- Released
- 2026-10-12
- License
- proprietary
- Context window
- 1M tokens
- Knowledge cutoff
- Not disclosed
- Verified
- source checked 2026-10-13
- Parameters
- Not disclosed
- Architecture
- Not disclosed
Muse Wonder’s verified record
Muse Wonder’s most-compared rival is Claude Fable 5.1: 1 trail and 1 not callable across their 2 shared comparisons. Muse Wonder is priced at $4.00/$20.00 per 1M tokens (in/out) vs Claude Fable 5.1’s $10.00/$50.00.
Against the 53 head-to-head comparisons Muse Wonder shares with other tracked models:
A gap counts for Muse Wonder only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Muse Wonder trails on 7 of them.
HLE · no tools Reasoning
±2 is noiseNo verdict for Muse Wonder anywhere on Terminal-Bench 4.0 · max ( nothing independently confirmed on both sides).
- None of Muse Wonder’s agentic comparisons are independently confirmed on both sides yet.
Muse Wonder API pricing
$4.00 in / $20.00 out per 1M tokens — official pricing source
What Muse Wonder costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.600 |
| A codebase review | 1,000K / 100K | $6.00 |
| A day of agent work | 10,000K / 1,000K | $60.00 |
Computed from Muse Wonder’s list rates above — cache discounts and batch tiers are not applied.
Muse Wonder is one of 3 Meta (Meta Superintelligence Labs) models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| Muse Wonder | $4.00 | $20.00 |
| Muse Spark 1.3 | $1.25 | $4.25 |
| Muse Spark 1.2(previous version) | $1.25 | $4.25 |
Muse Wonder benchmark scores
| Benchmark | Score |
|---|---|
| HLE(no tools)[1] Reasoning · ±2 is noise | |
| Terminal-Bench 4.0(max)[2] Terminal ops · ±12.4 is noise |
Who ran these numbers: 2 of 2 independent — artificialanalysis.ai (2).
- HLE: Artificial Analysis model page, embedded data, effort slug max, checked 2026-10-13. Underlying hle 0.500926784059314, recorded as 50.09. Muse Spark 1.3 (Max) on the same payload is 0.487025023169602; this site stores that row as 48.7. The difference against the stored figure is 1.39, inside Humanity's Last Exam's 2.0 band.
- Terminal-Bench 4.0: Artificial Analysis model page, embedded data, effort slug max, checked 2026-10-13. Underlying terminalBench40 0.52020202020202, recorded as 52.02. Muse Spark 1.3 (Max) on the same payload is 0.333333333333333 (33.33). The difference is 18.69, past Terminal-Bench 4.0's 12.4 band. This site's recorded Spark row stays vals.ai's 24.75, a different runner, so that stored pair is not this comparison.
Notes on the record
Launched 2026-10-12. Meta's model card lists a 1M context, which Artificial Analysis records as 1,048,576 tokens, and one rate of $4.00 in / $20.00 out per million tokens, cached input $0.20 (dev.meta.ai/models/muse-wonder and the research post, both checked 2026-10-13). No length threshold appears on that card. Meta publishes no parameter count and no knowledge cutoff; Artificial Analysis leaves both null on the max-effort row. No weights are linked, so the licence field is proprietary.
Two tracked benchmarks have an Artificial Analysis max-effort number, read from that page's embedded data the same day. Humanity's Last Exam without tools is 50.09. Muse Spark 1.3's max-effort figure on the same payload is 48.70, and this site stores 48.7; the gap is 1.39 points, inside the 2.0 band, so the order is a tie. Terminal-Bench 4.0 is 52.02 against Muse Spark 1.3's 33.33 on that same max-effort row, a gap of 18.69 points, past the 12.4 band. This site's recorded Spark row is vals.ai's 24.75, a different runner, and is not that comparison.
Muse Wonder Full shipped the same day, only inside Muse Code's Power and Maximum plans, with no per-token price. Artificial Analysis scores a max-effort row for it: Humanity's Last Exam 53.61 and Terminal-Bench 4.0 57.07, and the price fields on that object are null. It is not a separate record here.
On the same Muse Wonder object, GPQA, LiveCodeBench, Terminal-Bench 2.1 and AA-AnalystAgent are null. That describes Artificial Analysis's page on 2026-10-13, not every other board.
Compare with
FAQ
What does Muse Wonder cost?
The card price is $4.00 per million input tokens and $20.00 per million output tokens, cached input $0.20, checked 2026-10-13. One rate covers the 1,048,576-token window; the card shows no prompt-length step. Muse Spark 1.3's Standard tier is $1.25/$4.25 with a $0.15 cache read, so Muse Wonder's list price is higher on input, output and cache.
How does Muse Wonder compare with Muse Spark 1.3?
On Artificial Analysis at max effort, checked 2026-10-13, Humanity's Last Exam is 50.09 against Muse Spark 1.3's 48.7. That 1.39-point gap sits inside the 2.0 band, so the order is a tie. Terminal-Bench 4.0 is 52.02 against Spark's 33.33 on the same runner, 18.69 points, past the 12.4 band. Spark's recorded row on this site is vals.ai's 24.75 and used a different runner.
What has Muse Wonder been benchmarked on?
Two independent scores, both from Artificial Analysis at max effort on 2026-10-13: Humanity's Last Exam 50.09 without tools, and Terminal-Bench 4.0 52.02. GPQA, LiveCodeBench, Terminal-Bench 2.1 and AA-AnalystAgent are null on that same object. Both of Muse Wonder's two scores are independent runs.
Is Muse Wonder good at terminal work?
On the terminal benchmark Artificial Analysis has scored, Muse Wonder's 52.02 beats Muse Spark 1.3's 33.33 at matched max effort, and 18.69 points clears Terminal-Bench 4.0's 12.4-point band. That is one pair on one runner. It does not place Muse Wonder in the rest of the column, and Terminal-Bench 2.1 is null on the same page.
Can I call Muse Wonder Full from the API?
Not at a published token rate. Meta's developer post puts Muse Wonder on the Meta Model API at $4/$20, and limits Muse Wonder Full to Muse Code's Power Usage and Maximum Usage plans. Artificial Analysis still has a max-effort row for Full: Humanity's Last Exam 53.61 and Terminal-Bench 4.0 57.07, with no input price on that object.
Is Muse Wonder open source?
No. Meta's model card does not link a checkpoint, and Artificial Analysis lists the max-effort row as proprietary with a null parameter count. The research post does not publish a knowledge cutoff either; Artificial Analysis leaves that field null.
Further reading
- Models with 10M token context windows 2026 — Muse Wonder is one of the 38 models it compares.
Benchmark guides
- Humanity's Last Exam — what Muse Wonder’s reasoning score on it does and doesn’t prove.
- Terminal-Bench 4.0 — what Muse Wonder’s agentic score on it does and doesn’t prove.