Meta Launches Muse Wonder at $4/$20 MTok, Keeps the Stronger Wonder Full Off Its API

TL;DR

Meta released Muse Wonder and Muse Wonder Full on October 12, its first models that credibly sit at the frontier rather than a discount behind it. Artificial Analysis puts Wonder Full at 56 on its Intelligence Index, two points under Claude Opus 5.5 and three above GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon. Meta's own table has Wonder Full leading on SWE-Marathon, DeepSWE 1.1, MemoryBench and legal agent work. The catch: Wonder Full is not on the API. The model you can buy per token is plain Muse Wonder (index 52), priced at $4 input, $0.20 cached, $20 output per million tokens, a line-for-line copy of Opus 5.5's price list. Wonder Full ships only inside the $50 and $200 Muse Code subscriptions, with no per-token price and no self-hosting.

What shipped

Both models come out of Meta Superintelligence Labs, the group Alexandr Wang has run since the August 2025 restructure. They were trained at Meta's gigawatt-scale Prometheus data center in central Ohio, through pretraining and reinforcement learning, and Meta calls them "the next step on our scaling ladder." Both take text, image, video and audio in, text out, with a 1M-token context window, unchanged from Muse Spark 1.3.

Two product features get the marketing weight. "Active Memory" carries context across conversations and retrieves details when they matter. "Active Focus" keeps an agent working toward a goal, tries alternative approaches when one fails, and checks in before consequential actions. The launch demos are consumer-agent fare: a 3D kitchen game, a running plan assembled across apps, swag ordered off Instagram photos, a wedding seating chart.

Per VentureBeat, Zuckerberg described Wonder as "a much larger and more advanced class of models" than Muse Spark, which was built for low latency. Meta says a full safety report with more evaluations is "available soon." The weights are closed.

The two-tier catch

Read the developer page carefully, because the headline numbers and the price belong to different models.

• Muse Wonder (muse-wonder-1.0): on the Meta Model API and in Muse Code. $4.00 input, $0.20 cached input, $20.00 output per million tokens, flat across the whole 1M context. VentureBeat reports limits of 3,000 requests and 4 million tokens per minute, and that the endpoint speaks the OpenAI- and Anthropic-style tool interfaces.
• Muse Wonder Full (muse-wonder-full-1.0): Muse Code subscribers only, on the $50/month Power Usage plan (20x the usage of the $5 Everyday plan) and the $200/month Maximum Usage plan (80x). No per-token price, no API.

One Hacker News reader who went to look up pricing noticed that the benchmark scores on the main marketing page are Wonder Full's, while the price shown is Wonder's. The benchmark-topping model and the one you can buy per token are not the same model, which is a neat trick if nobody reads the footnotes.

VentureBeat also reports the API runs only in "standard" mode, where submissions are not used for training. Muse Code subscriptions default to "contributor" mode, which permits Meta to use prompts, code, files and outputs to improve models. You can switch it off, but only for future submissions, and Meta's terms bar confidential or personal data in that mode. Wonder Full, the model you would most want for serious work, is reachable only through the product whose default is to train on your work.

The independent numbers

Artificial Analysis ran both models at max reasoning effort on its Intelligence Index v4.3 the day of launch. Wonder Full scored 56 and Wonder 52, against 48 for Muse Spark 1.3. Opus 5.5 holds the top at 58, with Claude Sonnet 5.5 tied with Wonder Full at 56 and GPT-6 Astra, Fable 5.1 and Gemini 4 Argon clustered at 53.

The sub-scores are where it gets interesting. On Terminal-Bench 4.0, Wonder Full went from Spark 1.3's 33% to 57%, and The Decoder notes it did so while using about 30% fewer output tokens. On the Coding Agents Index, Wonder Full inside Muse Code scores 68, tied with Claude Code running Sonnet 5.5 and above Claude Code with Opus 5.5 at 66; Muse Code with Spark 1.3 had scored 54. Wonder Full took narrow first places on CritPt and GDP.pdf (33%, with the nearest competitors at 31-32%) and placed second on GDPval-AA at 1857 Elo, within the confidence intervals of Opus 5.5 (1866) and Sonnet 5.5 (1838).

Knowledge is the soft spot. On AA-Omniscience, Wonder Full scores 30 and Wonder 25, against 46 for Opus 5.5, 43 for Astra and 42 for Argon. The hallucination rate did improve, from 33% on Spark 1.3 to 23% on Wonder Full (27% on Wonder), second-lowest among models scoring 45 or more. Gemini 4 Argon is lowest at 15%. So it is a model that admits ignorance more gracefully than its predecessor, while knowing noticeably less than its rivals.

Meta's own table

The developer page carries a comparison table Meta compiled from internal evals, vendor-supplied scores and third-party results, which VentureBeat rightly flags as a mixed bag. Taken at face value, Wonder Full leads on the agentic coding and professional rows: SWE-Marathon (long-running coding) 57.5 versus 46.3 for Opus 5.5 and 42.5 for Astra; DeepSWE 1.1 at 79.0 versus 74.2 and 74.1; Harvey's LAB-AA legal benchmark at 13.1 versus 4.2 and 8.6; MemoryBench at 85.5 versus 81.5 and 81.4 (plain Wonder scores 85.6 there, a tenth higher than Full).

The same table shows where Meta loses. Opus 5.5 and Astra both beat Wonder Full on Terminal-Bench 4.0 (59.6 and 59.1 versus 57.1), on HealthBench Pro (65.6 and 65.3 versus 62.3), and Opus 5.5 is far ahead on ProgramBench, 70.5 to 59.0. Opus 5.5 also edges JobBench, 75.4 to 75.3. Meta published the losses in its own table, which is more than most launch pages manage, and is a reasonable sign the wins are not pure cherry-picking.

One early hands-on data point cuts the other way. A Hacker News user ran an image-to-HTML test on plain Wonder and found it "much worse than Opus 5.5," closer to Sonnet or Haiku, and guessed the vision side is less mature than the text side. One test, one user, but it matches the shape of the AA results: strong on long agentic text loops, weaker everywhere else.

The price is the story

Meta's previous Muse Spark releases came in cheap: Spark 1.3 lists at $1.25 input and $4.25 output. Wonder is $4 and $20, a 3.2x and 4.7x jump, and exactly what Anthropic charges for Opus 5.5 ($4 / $0.20 cached / $20). It undercuts Astra and Fable 5.1, which both list at $10 / $50. As one HN commenter put it, that is a signal Meta now believes it has reached the frontier and will price like it, rather than buying users with discounts on a slightly worse model.

Per-token price is not what you pay, though. Artificial Analysis estimates the cost to run one of its Intelligence Index tasks, which folds in how many tokens each model burns to get there. A per-token price is like the price per litre at the pump; cost per task is the fuel bill for the whole trip, and Wonder is a thirsty car. It consumed 7.69M input tokens per task against 6.03M for Spark 1.3, and The Decoder calculates input makes up 76% of its cost.

So on the API you are paying Opus 5.5 rates for a model that scores 52, when Astra scores 53 at $3.26 a task and Argon scores 53 at $1.99. The value argument for Wonder on the API is the agentic sub-scores, not the headline index. The value argument for Wonder Full is that $50 a month of Muse Code buys a 56-point model; Anthropic's comparable plans sell a 58-point one, and you will have to decide what SWE-Marathon is worth to you.

Caveats

• Leaderboard position moved within hours. HN readers saw Wonder Full at second place on Artificial Analysis in the morning and fifth, behind Gemini 4 Argon, later the same day. The AA model page lists it at #4 of 699 at the time of writing. Treat the scores as settled and the rank as not.
• Meta's table is vendor-compiled. It mixes internal runs, competitor-published numbers and third-party results, and some coding comparisons used different tools. The Artificial Analysis figures are the independent ones.
• Wonder Full has no cost per task. No per-token price means no way to compare it on the AA cost chart, and no way to know what it would cost if Meta ever meters it.
• Muse Code is the only door to Wonder Full, and HN complaints about its sandbox failing on stock Mac and Linux installs, and about the $5 Everyday plan burning 14% of a weekly allowance on three small requests, predate this launch. One commenter also reports subscriptions remain North America only.
• Everything here is text-out, closed-weight, Meta-hosted. No self-hosting, no quantized GGUF, no llama.cpp. Not the Meta of 2024.

Key Takeaways

• Meta shipped Muse Wonder and Muse Wonder Full on October 12; Artificial Analysis scores them 52 and 56, versus 48 for Muse Spark 1.3 and 58 for Claude Opus 5.5.
• Only Wonder is on the API, at $4 / $0.20 / $20 per million tokens, identical to Opus 5.5. Wonder Full is locked to the $50 and $200 Muse Code plans.
• Meta's own table has Wonder Full ahead on SWE-Marathon (57.5 vs 46.3 for Opus 5.5), DeepSWE 1.1, MemoryBench and Harvey LAB, and behind on Terminal-Bench 4.0, HealthBench Pro and ProgramBench.
• Knowledge lags: AA-Omniscience 30 for Wonder Full against 46 for Opus 5.5, though hallucinations fell from 33% to 23%.
• Cost per AA task is $4.52 for Wonder, 2.8x Spark 1.3 and above GPT-6 Astra, which scores higher. Price Wonder on agentic sub-scores, not the index.
• Muse Code subscriptions default to contributor mode, which lets Meta train on your prompts and code; the API does not.

Sources: Meta, Muse Wonder announcement, Meta developer page: Muse Wonder pricing and benchmark table, Artificial Analysis, Muse Wonder and Muse Wonder Full analysis, Artificial Analysis, Muse Wonder Full model page, VentureBeat, TechCrunch, The Decoder, Hacker News discussion