Muse Wonder 1.0 Review: Features, Pricing, Pros and Cons

Meta priced Muse Wonder to the cent: $4 per million input tokens, $20 per million output, 20 cents per million cached. Those are exactly Claude Opus 5.5's list prices. Coincidence, or a statement of intent.

Then, on 27 October, the independent benchmarking firm Artificial Analysis rewrote its exam. Opus 5.5 scored 58. Muse Wonder scored 43.

Same price tag, fifteen points apart. That gap is why this Muse Wonder 1.0 review spends more time in the fine print than in the launch highlights.

A quick housekeeping note: frontier labs now ship new models every few weeks, and Muse Wonder is still officially in preview. Everything below reflects Meta's documentation and public benchmarks as of 30 October 2026, so check Meta's developer site before you commit a production budget.

Muse Wonder 1.0 is Meta's closed-weight frontier model for coding and long-running agents, released on 12 October 2026 and sold through Meta's Model API at $4/$20 per million tokens with a 1-million-token context window. Is it worth it? For multi-hour coding agents in real codebases, it deserves a serious test. For almost everything else, there's a cheaper or stronger option.

Muse Wonder 1.0 at a Glance

Sources: Meta's model card and developer docs. Note the "Weights" and "Sibling model" rows. Both matter.

What Muse Wonder Is, and Why It Exists

Muse Wonder is the top rung of a ladder Meta has been climbing fast. Meta Superintelligence Labs launched Muse Spark in April 2026, iterated to Spark 1.3, and shipped Wonder about six weeks later, trained at Meta's gigawatt-scale Prometheus data centre.

We covered the launch, and what it signals about Meta, in our news story on Muse Wonder. This piece answers the other question: should you actually use it?

Meta calls these its "most capable models" yet. True, and Artificial Analysis agrees: Wonder scored 52 on its launch-week index against 48 for Muse Spark 1.3. But that's Meta comparing Meta to Meta. Being the tallest in your own family says nothing about the neighbours.

Muse Wonder vs Muse Wonder Full: two models, not two settings

Meta launched two models on the same day, and they are not one model at two power levels.

• Muse Wonder (muse-wonder-1.0) is the one you can build on: on the Meta Model API, priced per token, and on every Muse Code plan.

• Muse Wonder Full (muse-wonder-full-1.0) is "the higher-capability variant." It lives only in Muse Code, Meta's coding tool, on the $50 and $200 plans. No API, no per-token price, no way into your own app.

• Turning Wonder up doesn't make it Full. Meta's docs say setting reasoning effort to "max" on Wonder does not select Full.

Think of a restaurant whose best dish is only served at the members' table upstairs. Everything else can be delivered. The dish on the poster can't.

Meta discloses no architectural differences between the two, but Full scores higher on most of Meta's own benchmarks. This review covers Muse Wonder 1.0, the API model. Where a number belongs to Full, we say so.

What's Actually New in Muse Wonder 1.0

Muse Wonder is built for agents, not small talk: long jobs that plan their own work, delegate to sub-agents, operate browsers and desktop apps, and edit many files in large codebases. What stands out:

• Active Memory and Active Focus. Meta's headline features: remembering relevant details across conversations, and pushing through hard tasks instead of quitting. Both are measured mainly on Meta's own internal tests, so treat them as promising, not proven.

• Four input types. Text, images, video and audio (plus PDFs) in, text out. Few leading models accept audio and video at all.

• Reasoning you can't turn off. Effort runs from "minimal" to "max", but "none" returns an error. Your bill will notice.

• A flat-priced 1M-token window. Same rate whether your context is nearly empty or nearly full.

• Easy swapping. Meta's API takes OpenAI-style requests, and moving between Muse Spark and Wonder is a one-line model ID change.

What isn't new is transparency: no parameter count, architecture, knowledge cutoff or training data details. Any parameter figure you see online didn't come from Meta.

Muse Wonder Benchmarks: What the Numbers Actually Say

Most launch coverage quoted the Muse Wonder Full column of Meta's benchmark table, because it's the flattering one. Here's the column for the model you can actually buy, plus the one most charts leave out: who produced each score.

What Meta reports (vendor-published model card, 12 October 2026, all at max effort)

Bold marks each row's best. Full beats plain Wonder on most rows, by 0.2 (DeepSWE) to 9 points (ProgramBench).

Now read the last column. On the two coding tests where every model ran in the same setup, Terminal-Bench and ProgramBench, Wonder ties for last and finishes last. Its biggest coding lead, SWE-Marathon, carries Meta's own footnote: Astra ran in OpenAI's Codex and Fable in Claude Code, so "this row is not a controlled common-harness comparison."

A harness is the scaffolding a model works inside: the tools, prompts and loop that let it edit files and run tests. Think of it as running shoes. A race where everyone wore their own shoes is partly a race between shoes.

Scepticism cuts both ways, though. On DeepSWE 1.1, Datacurve scored Wonder and Opus 5.5 under the same conditions, and Wonder wins 78.8 to 74.2, like for like. That's the strongest single point in its favour.

Even on Meta's own house index, the Personal Superintelligence Index, the API model trails Astra and Opus 5.5. Full is the one winning Meta's chart.

The fine print Meta did publish

To its credit, Meta's card is open about method and includes rows Wonder loses. Three details matter:

• Score picking. Where no outside result exists, Meta uses "the highest of our internally validated run, self-reported score, or official leaderboard score." Translation: its best number.

• "Best-effort" rival runs. Meta admits its setup "may not be optimally tuned for proprietary models."

• Same name, different ruler. Meta's table gives Wonder 39.1 on AutomationBench-AA under strict all-checks-must-pass scoring. Artificial Analysis' own write-up quoted 68% for the same model on the same benchmark. Check the scoring before you compare charts.

What independent testing shows

Eighteen days after launch, independent data exists, and it's less flattering.

Artificial Analysis Intelligence Index v4.3 (12 October 2026). In AA's launch-day evaluation, Wonder scored 52 and Full 56, against 58 for Opus 5.5 and 53 for both Astra and Fable 5.1. One point behind Astra. Respectable.

Artificial Analysis Intelligence Index v5.0 (27 October 2026). AA's new version runs 60% on private or held-out tests, adds Cognition's FrontierCode 2.0 and drops older tests such as SciCode. AA doesn't present the two versions as comparable, so don't read "52 to 43" as a fall. Read the standings.

On v5, Muse Wonder (max) scores 43, 23rd of 52 models, against 58 for Opus 5.5, 57 for Astra and 55 for GPT-6.1 Sol. The gap to Astra went from one point to 14. AA hasn't said why, and we won't guess. But when an exam goes private and one student drops further than the class, test on your own work before trusting anyone's chart. Full has no v5 score because it has no public API.

Coding Agent Index (model plus tool). At launch, AA scored Muse Code at 63 running Wonder and 68 running Full, above Claude Code on Opus 5.5 (66). On AA's revised v2.0 index, Claude Code on Opus 5.5 leads at 61.7, with Muse Code on Full third at 58.9. We found no v2.0 score for plain Wonder.

SWE-Marathon's public board. As reported on 13 October, it showed Wonder at 52.5% and Full at 53.1% in Muse Code, with Claude Sonnet 5.5, a $2/$10 model missing from Meta's table, at about 51%. On Meta's showcase test, Wonder's edge over a half-price model is about a point and a half.

Knowledge, honesty, speed. Wonder scored 25 on AA-Omniscience, a knowledge test, against 46 for Opus 5.5, yet its 27% hallucination rate was third-lowest among models AA compared at launch. It knows less, but bluffs less. Output runs at about 197 tokens a second, but time to first token is 23.6 seconds, against a 3.6-second median.

Muse Wonder Pricing: What It Really Costs

On the Meta Model API, Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, as of 30 October 2026.

All prices are per million tokens. Wonder is 60% cheaper per token than Astra or Claude Fable 5.1, and identical to Opus 5.5, itself a price cut on the $5/$25 model in our Claude Opus 5 review.

The sticker isn't the bill, though. Here's what actually moves it:

• Reasoning is always billed at the $20 output rate, and can't be switched off. Even "summarise this sentence" pays for some thinking.

• It's wordy. On AA's v5 index, Wonder produced 110 million output tokens against a 64 million median.

• Long context is flat-rate. GPT-6 Astra re-prices the whole request above 272,000 input tokens. Wonder doesn't, a genuine advantage for giant prompts.

• Caching is automatic, with cached input 95% off. Agents that resend long context benefit most.

• Web search is extra, at $2.50 per 1,000 searches. No batch discount is listed.

The real test is cost per task, which AA measures. On v5, Wonder averaged $6.95 per task, against $6.24 for Astra, $10.04 for Opus 5.5, $13.04 for Fable 5.1 and $1.27 for GPT-6.1 Sol. Sol scores 12 points higher for under a fifth of the cost.

Is Wonder worth it over Muse Spark 1.3?

For most readers, Wonder's real rival is its younger sibling. Muse Spark 1.3 costs $1.25/$4.25. On AA's launch-week index it scored 48 to Wonder's 52, at $1.60 per task against $4.52. That's 2.8 times the cost for four points.

Meta's own docs agree: Wonder where "capability matters more than cost per token," Spark for "high-volume, latency-sensitive, or cost-sensitive work." Run Spark by default and route the long, hard jobs to Wonder.

Muse Code: the subscription route

A $5 plan is a cheap way to kick the tyres. Two catches: it isn't API credit, and every plan starts in Contributor mode, which lets Meta train on your work.

Can you self-host it to save money?

No. Wonder has no downloadable weights, so the deployment economics are simple: you pay Meta's meter on Meta's servers.

Meta's open option is a different model, Muse Spark Open, released on 15 October 2026: 816 billion parameters, 42 billion active per token. Its weights alone take about 537GB compressed, served on four Nvidia B300s per Meta's docs, and it's weaker, at 39 on AA's older v4.3 index. For open-weight versus closed-API cost maths, see our DeepSeek V4 vs Claude Fable 5.1 comparison.

Is Muse Wonder Open Source? The Licence, Read Properly

No. Muse Wonder and Muse Wonder Full are closed-weight, so there's no model licence in the Llama sense. You're buying a service under the Meta Model API Terms of Service (updated 9 October 2026), Meta's Acceptable Use Policy and Muse's supplemental terms.

Meta's old Llama licences carried a 700-million-monthly-active-user threshold. The Wonder terms have no user or revenue threshold at all. These clauses matter more:

• You own your outputs (Section 4.1), to the extent the law allows.

• No distilling. Section 10.1(ix) bars using the service or its outputs to "train, fine-tune, distill, or otherwise improve or create" a competing model without written permission. Evaluating your own systems is fine.

• No benchmarking to promote a rival. Section 10.1(xi) bars publishing benchmarks of the service to promote a competing product without permission.

• Adults only. You and your end users must be 18 or older.

• Two data modes. The API runs in Standard mode, which Meta says doesn't train on your content. Contributor mode, Muse Code's default, does, and Section 6.2 says not to submit sensitive, confidential or personal information there.

• Preview-grade guarantees. Meta can change features or suspend access "at any time without notice," update fees with no stated notice period, and caps its liability at the greater of 12 months' fees or $1,000.

Put the default next to Section 6.2 and the problem jumps out: Muse Code starts you in a mode where Meta's own terms say not to submit confidential material. That describes most client codebases.

The nearest Llama-style condition sits on the open model instead. Muse Spark Open's licence allows commercial use, requires a "Built with Muse Spark Open" label, and makes model-as-a-service providers above $20 million in annual revenue sign a separate deal with Meta.

Where Muse Wonder Is Strong, and Where It Falls Short

Where it's strong:

• Long software-engineering tasks. It beats Opus 5.5 on DeepSWE 1.1, like for like.

• Legal agent work, relatively: 10.0 on Harvey LAB-AA leads all three rivals, though that's about one task in ten fully done.

• Huge prompts at a flat price across the whole 1M window.

• Audio, video, image and PDF input in one model.

• A low 27% hallucination rate, and fast output once it starts.

Where it falls short:

• It isn't Meta's best model, and the best one isn't on the API.

• It's mid-pack or last on coding tests where every model shared a setup.

• It ranks 23rd of 52 on AA's v5 index, 14 points behind Astra.

• A 23.6-second wait for the first token rules out live chat.

• Closed weights, and no AWS, Azure or Google Cloud listing that we could find as of 30 October.

• A thin knowledge score: 25 on AA-Omniscience against 46 for Opus 5.5.

• Meta's card flags shakier multi-step reasoning in unfamiliar situations, weak 3D and spatial reasoning, and lower quality outside its strongly supported languages.

• The full safety report promised at launch is still "available soon," and Meta corrected one cyber-security result on 23 October after fixing a testing error.

How to Actually Get Muse Wonder

1. Meta Model API. Get a key on Meta's developer platform, point an OpenAI-style client at Meta's endpoint, and call muse-wonder-1.0. Limits are 3,000 requests and 4 million tokens per minute per team, in Standard mode only.

2. Muse Code, from $5 a month for Wonder or $50 for Full. Switch off Contributor mode in /settings first. Per Meta's subscription docs, the setting is saved per computer, so repeat it on each machine, and use --contributor-mode off for scripted runs. Meta only promises that future content is excluded.

3. OpenRouter, as meta/muse-wonder, though not through its EU in-region routing.

Wonder isn't available in South Korea, and Meta's Geographic Use Policy restricts other countries too.

Who Should Use Muse Wonder, and Who Should Skip It

Use it if:

• You run multi-hour coding agents on existing codebases and can test Wonder against your current model on your own repo.

• You regularly send prompts well past 272,000 tokens.

• You already build on Meta's API and want a step up from Spark for the hardest jobs.

Skip it if:

• You want capability per dollar. GPT-6.1 Sol beats it on both counts.

• Your users sit watching a cursor blink, as in live chat.

• Your tasks are high-volume and simple. Spark 1.3 does that for a fraction of the price.

• You need weights, on-premise hosting or your cloud provider's marketplace.

• You hoped to distil a frontier model into your own. The terms forbid it.

• You just want a better everyday assistant. Our ChatGPT vs Claude comparison is the more useful read.

Muse Wonder 1.0 Review: The Verdict

Muse Wonder 1.0 is Meta's most serious model in years, and on long, messy coding jobs it's genuinely competitive. The DeepSWE win is real.

But the model you can buy isn't the model on the poster. Meta's chart leans on Full, Wonder's biggest coding leads come from mismatched setups, and the newest independent index puts it 23rd. It's priced like Opus 5.5 without performing like it.

So don't make Wonder your default yet. Test it on your own long-running coding tasks, through the API or a $5 Muse Code plan with Contributor mode off. If it wins on your repo, route your hardest jobs to it. If it doesn't, you've spent five dollars learning something no chart could tell you.

Meta finally has a frontier model worth testing. It just kept the best one for the members' table.