Meta released two models on Monday 12 October, Muse Wonder and Muse Wonder Full, and every headline number in the launch belongs to the second one. Artificial Analysis scored Full at 56 on its Intelligence Index, two points behind Claude Opus 5.5 and level with Sonnet 5.5. Wonder scored 52, one point under GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1.

Only Wonder is in the Meta Model API, as muse-wonder-1.0, at $4 per million input tokens and $20 per million output, the same as Opus 5.5 to the cent. Full is listed as "Not available" on the API and is sold only inside Muse Code, on the $50 Power Usage and $200 Maximum Usage plans. Meta's models page says the rest in one sentence: "setting reasoning effort to "max" on Muse Wonder doesn't select it."

At the Opus price, Wonder costs $4.52 per Intelligence Index task against $5.98 for Opus 5.5 and $3.26 for Astra, and 76% of that bill is input, 7.69 million tokens per task. The subscription that carries Full defaults to Contributor mode, where your prompts, code and files train future Meta models, and its key "is for use with Muse Code only."

What Meta shipped

Meta Superintelligence Labs published the launch on its research blog and in the company newsroom on 12 October, without a byline. The post calls the pair "the next step on our scaling ladder and a major leap toward personal superintelligence" and says both "build on larger, stronger base models" than the Muse Spark line. The model card gives them a 1M-token context, input in text, image, video and audio, text out, and nothing about how the two differ in size, training or serving. The models page calls Full "the higher-capability variant" and leaves it there.

Availability is two sentences on the blog. "Muse Wonder is available today through the Meta Model API and Muse Code." "Muse Wonder Full is available today for Muse Code subscribers." The consumer Muse assistant gets Wonder "soon", subscribers first. TechCrunch reported that both were trained at Meta's Prometheus data center in central Ohio.

The Hacker News thread reached 669 points and 380 comments, and the recurring question in it is which model people were reading about. One commenter put it flatly: "Only Muse Wonder is available in the API." Another: "they're not even making Wonder Full available outside of their own coding harness." Simon Willison listed the models his API key could see, and Full was not among them. It was not on OpenRouter either.

Two models, one scorecard

The model card runs both Meta models at max reasoning effort against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, also at max. The rows where the two Meta models land apart tell you what the API is selling.

On the Artificial Analysis Intelligence Index, Full scores 56 and Wonder 52. On Terminal-Bench 4.0, 57.1 against 52.0. JobBench, 75.3 against 70.5. AutomationBench-AA, 45.4 against 39.1. OSWorld V2.1, 78.4 against 72.6. ProgramBench, 59.0 against 50.0, with Opus 5.5 at 70.5. The two rows where the gap vanishes are DeepSWE v1.1, 79.0 against 78.8, and MemoryBench, Meta's internal recall benchmark, where Wonder edges Full 85.6 to 85.5 and both beat Opus at 81.5.

Counted across the card's fourteen rows, Opus 5.5 beats the API model on nine and beats Full on six. Full beats Astra on eleven, and Wonder loses to Astra on eight. The card's own methodology notes say how loosely to hold the rest. Where no third-party or provider result exists, Meta uses "the highest of our internally validated run, self-reported score, or official leaderboard score." Competitor runs are "best-effort: prompts, tools, and runtime may not be optimally tuned for proprietary models." The SWE-Marathon row, where Full posts 57.5 to Astra's 42.5, carries the note that it "is not a controlled common-harness comparison" and that Full "was reported with xhigh reasoning effort", not max.

Artificial Analysis published its own runs the same day. Full is first on CritPt and GDP.pdf at 33%, second on GDPval-AA at 1857 Elo to Opus 5.5's 1866, and well back on AA-Omniscience at 30 to Opus's 46. On the Coding Agents Index, Muse Code running Full scores 68, level with Claude Code on Sonnet 5.5, and running Wonder it scores 63. Six weeks earlier Muse Spark 1.3 scored 48 on the Intelligence Index. The card's risk section is Full only too: "Results below are from Muse Wonder Full, our most capable variant, which is why the assessment covers every model in the family."

What the API model costs per task

Wonder's price card is Opus 5.5's: $4 input, $0.20 cached input, $20 output per million tokens, with reasoning tokens billed as output. That is double Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon's introductory rate, and 60% below Astra and Fable 5.1 at $10 and $50. Meta's pricing page adds a line the others do not: "No long-context premium: you pay the same rate whether your context window is mostly empty or almost full." Web search grounding is $2.50 per 1,000 queries on top. Rate limits are 3,000 requests and 4 million tokens per minute, per team rather than per key, and Wonder is still marked "in preview".

Artificial Analysis measured $4.52 per Intelligence Index task for Wonder at max effort. Fable 5.1 costs $7.63 per task, Opus 5.5 $5.98, Sonnet 5.5 $5.46, Astra $3.26, Argon $1.99, Sol $0.72. The split is the part to keep. $3.43 of Wonder's $4.52 is input, from 7.69 million input tokens per task: $1.92 in cache writes, $1.44 in cache hits, $0.07 uncached. Output is about 54,000 tokens per task and $1.08. Wonder is not a model that writes a lot. It is a model that reads a lot, and at $4 a million the reading is the bill. Full has no cost-per-task figure because, as Artificial Analysis wrote, it "is not priced per token".

The reasoning page is where the API gotcha lives. reasoning_effort takes none, minimal, low, medium, high, xhigh and max, and none returns a 400 on Wonder. The models page says of Full: "Muse Wonder Full, the higher-capability variant, is available only in Muse Code with a subscription, so setting reasoning effort to "max" on Muse Wonder doesn't select it." Max buys more reasoning tokens from the same model at the same price per token, and max_tokens caps reasoning and visible output together, so a long think can truncate the answer you paid for.

The plan that has Full

Muse Code sells four plans. Everyday Usage is $5 a month for Muse Spark, Muse Wonder and Muse Image, with "10-50 prompts every 5 hours". High Usage is $15 for five times that. Power Usage is $50 and is the first tier with "Access to the most powerful Muse models, including Muse Wonder Full", at 20 times Everyday's allowance. Maximum Usage is $200 at 80 times. The models page draws the same line: Full on "Power Usage and Maximum Usage", Wonder on "every plan".

Two sentences on the same help page set the terms. "Muse Code subscriptions default to Contributor mode", in which your prompts, code, files and the responses you get back "may be used to train future Meta models" in exchange for more usage, and "you can turn it off at any time." The subscription "applies to the Muse Code API key", a credential that "is for use with Muse Code only." The pay-as-you-go Wonder runs in Standard mode only, with Contributor "Not available". So the one model in the family that beats Astra on Meta's own card is reachable through a terminal agent whose default is to train on what you feed it, and not through the key your application uses. That is the Contributor tier that launched with Muse Code in August, where the prompt is whatever the retrieval step pulled into the 1M window.

The thread has the first quota reports. One $5 subscriber "blew through 14% weekly usage" on simple tasks, and another found two small prompts to "Muse Wonder Max-contributor used 4% of my weekly quota." The quality verdicts ran from "Intelligence is there I think, but 'taste' and 'judgment' is not there yet" to "It's much worse than Opus 5.5," and neither says which of the two models was running, which is the point of this piece. Open weights do not appear in Meta's text at all, and the Spark weights promised in September have not shipped either.

What to do

Benchmark your own prompts on muse-wonder-1.0, the model you can buy, and treat any number next to the words Wonder Full as a different product. Watch cached input before output length: at a 76% input share, your cache hit rate moves the bill more than anything the model writes. Set max_tokens with reasoning included, and do not expect max to change which model answers. If you want Full, budget a $50 Power plan per developer, switch Contributor mode off before the first prompt touches client code, and accept that it is not callable from your pipeline. And if Meta's card is your reference, read its three methodology bullets before the table.