Cowry Tech News
AI Cloud

Meta’s Muse Wonder is here. Builders get the API model. The flagship stays inside Muse Code.

Meta launched Muse Wonder and Muse Wonder Full on October 12, calling them its first frontier models. Three days on, the shape is clear: Muse Wonder is on the Meta Model API at $4/$20 per million tokens, while the stronger Full model is locked to paid Muse Code plans. Independent tests put both near the top of the field, but they don't fully match Meta's own benchmark story.

By Cole Harlan · 15 Oct 2026 · 13 min

Meta developer launch artwork reading Build with Muse, available on Meta Model API
Image: Meta

Meta released two new AI models on Monday, October 12: Muse Wonder and Muse Wonder Full. Meta’s research blog calls them “our most capable models that deliver state-of-the-art performance on personal and professional agentic tasks.” TechCrunch, separately, framed the pair as Meta’s first attempt at a frontier model.

This story comes three days after the launch. We held it to check Meta’s pricing pages, the model card and the first independent evaluations against what Meta says. The result is more mixed than the launch posts.

The short version for builders: you can call Muse Wonder through an API today. You cannot call Muse Wonder Full through an API at all. Full, the model Meta says leads coding leaderboards, is only available inside Meta’s Muse Code subscription.

Below we cover what each model is, where you can get it, what it costs against rivals, which benchmark claims come from Meta and which come from outside testers, and what Meta still hasn’t said.

What Meta launched

Meta’s developer blog post, written by Josh Walters, announced Muse Wonder “on Muse Code and Meta Model API at $4 input / $20 output per million tokens (USD).” Meta’s Model API pricing page lists the same Standard-mode rates for muse-wonder-1.0: $4.00 input, $0.20 cached input and $20.00 output per million tokens. The developer post also said Muse Wonder Full is “available exclusively by subscription in Muse Code.”

Meta describes the Wonder series as built for “complex coding and agentic work.” Its developer post lists multi-step reasoning, end-to-end agentic task completion, coding agents, tool use and multimodal reasoning. Its product card for Muse Wonder says the model “plans long tasks, delegates to sub-agents, and recovers from errors.”

The model card, published on Meta’s research site, lists the basics:

Item Muse Wonder / Muse Wonder Full (per Meta model card)
Release date October 12, 2026
Provider Meta Platforms, Inc. or Meta Platforms Ireland Limited
Input Text, image, video, audio
Output Text
Context length 1M tokens at inference
Intended use Commercial and research use

Artificial Analysis, an independent benchmarking firm, lists the same modalities and the same 1M-token context window. It notes the context window is unchanged from Meta’s earlier Muse Spark 1.3.

One caveat on audio. The model card lists audio as an input. But Meta’s developer post adds a line under multimodal reasoning: “Audio-dependent use cases require confirmation of audio support on the release surface.” In plain terms, teams should check that audio input actually works on the API or product they plan to use before they build on it.

Two models, two ways in

Muse Wonder is the one most developers can touch. It is on the Meta Model API and in Muse Code. Meta prices it per token.

Muse Wonder Full is not on the API. Meta’s developer post says it is available “exclusively through the Power Usage and Maximum Usage Muse Code subscription plans.” Meta calls it “our most advanced model yet.” TechCrunch and Artificial Analysis both confirm that Full is limited to Muse Code subscribers and not served through the API.

Meta also introduced a new Maximum Usage plan for Muse Code. Meta says it includes 80 times the Everyday Usage plan’s limit, with “the same five-hour and weekly resets.” Full is available on both the Power and Maximum plans.

Meta did not list subscription prices for the Power or Maximum plans in the posts we reviewed. We have not confirmed them, so we are not reporting a figure.

What that means in practice: a company that wants Meta’s best model inside its own product, pipeline or agent stack cannot get it today. It can get Muse Wonder. Meta says Muse Wonder “brings comparable performance to Muse Wonder Full.” The independent numbers below show a real gap.

Meta also plans to use the models beyond developer tools. Its research blog says: “Muse Wonder will soon power experiences in Muse, and will roll out first to subscribers.” Meta has not set a public date, so we treat the agent rollout as announced but not yet shipped. The blog names Muse Wonder for that rollout, not Muse Wonder Full.

What it costs

Muse Wonder’s API price, from Meta’s Model API pricing docs (muse-wonder-1.0, Standard mode):

Muse Wonder (Meta Model API) Price per 1M tokens
Input $4.00
Cached input $0.20
Output $20.00

Muse Wonder Full has no per-token price. Meta lists it as “included with a Muse Code subscription.”

How does that compare? Artificial Analysis says the $4/$0.20/$20 rate matches Claude Opus 5.5. It says Muse Wonder costs double each of Claude Sonnet 5.5, Gemini 4 Argon and GPT-6.1 Sol, which it lists at $2/$10. And it says Muse Wonder is 60% below GPT-6 Astra and Claude Fable 5.1, which it lists at $10/$50. We have not checked each rival’s own price page for this story, so treat those comparisons as reported by Artificial Analysis.

TechCrunch put it more loosely, calling Wonder “roughly equivalent in cost to GPT-6 Sol or Opus 5.5.” The Opus comparison lines up with Artificial Analysis. The Sol comparison does not line up with Artificial Analysis’s $2/$10 figure for GPT-6.1 Sol. Model names and versions here are close together, so buyers should check current list prices directly.

The jump from Meta’s last model is steep. Artificial Analysis says Muse Spark 1.3 was priced at $1.25 input, $0.15 cached and $4.25 output. Muse Wonder’s output price is more than four times that.

List price is only part of the bill. Artificial Analysis measured what it costs to run each model across its Intelligence Index tasks. Muse Wonder came in at $4.52 per task. That is lower than Claude Fable 5.1 ($7.63), Claude Opus 5.5 ($5.98) and Claude Sonnet 5.5 ($5.46), by its count. It is higher than GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99), GPT-6.1 Sol ($0.72) and Muse Spark 1.3 ($1.60).

The reason is input. Artificial Analysis found Muse Wonder used about 7.69 million input tokens per task, 28% more than Muse Spark 1.3 and two to seven times some rivals. Input made up about 76% of its per-task cost. Output use was moderate, at about 54,000 tokens per task. For builders, that points at caching. With cached input at $0.20 against $4.00 fresh, prompt structure will drive real costs on long agent runs.

Meta’s benchmarks

Meta’s claims are strong. Its developer post says Muse Wonder Full “tops coding benchmark leaderboards, including DeepSWE and SWE-Marathon in independent third-party runs, outperforming every other model evaluated.” Meta’s model card and developer comparison table put Wonder and Wonder Full against Claude Fable 5.1 and GPT-6 Astra; we do not reprint Meta’s individual scores because the row labels did not come through cleanly.

The model card explains how Meta built its comparison table, and the details matter:

  • Meta ran its own models at max reasoning effort and compared them with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 at max effort.
  • Where a third-party or provider-reported score existed, Meta used it. Otherwise it used the highest of its own validated run, a self-reported score or an official leaderboard score.
  • Meta says its internal runs of competitor models “remain best-effort” and that prompts, tools and runtime “may not be optimally tuned for proprietary models.”
  • For DeepSWE, Meta used Datacurve’s third-party results for its two models and Opus 5.5, OpenAI’s own reported figure for GPT-6 Astra, and Anthropic’s system-card figure for Fable 5.1.
  • For SWE-Marathon, Meta says results for GPT-6 Astra and Claude Fable 5.1 came from Codex and Claude Code, so the row “is not a controlled common-harness comparison.” Muse Wonder Full was reported at xhigh effort.

So the DeepSWE and SWE-Marathon claims rest partly on third-party runs. But the rows mix sources and harnesses. Meta says so itself.

Meta also leaned on memory. TechCrunch reported that Meta highlighted MemoryBench and said Muse Wonder scores higher on it than Fable, Opus or Astra. The model card describes MemoryBench as “an internal benchmark.” Several others in Meta’s agent suite are internal too, including its Personal Superintelligence Index, FocusBench and the Verifiable Browsing Task. These are Meta’s tests, run by Meta. We label those results vendor-reported.

Meta’s developer post includes a full score table. In the text we could retrieve, the benchmark row labels did not come through cleanly. We could not match every number to its benchmark with confidence, so we are not reprinting individual scores from Meta’s table.

What independent testers found

Artificial Analysis published results on launch day. Its headline: Muse Wonder Full scored 56 and Muse Wonder scored 52 on its Intelligence Index v4.3, both at max reasoning effort. Muse Spark 1.3 scored 48.

For context, when Artificial Analysis launched v4.3, it reported Claude Fable 5.1 and GPT-6 Astra tied at 53 at the top. On that basis, Full lands above both and Muse Wonder lands just below them. Its model page ranks Full fourth and Muse Wonder ninth on its leaderboard, and those ranks may move as new models are added.

Other findings from Artificial Analysis:

  • Full placed in the top three on six of the index’s ten evaluations. It ranked first on CritPt (33%) and GDP.pdf (33%), though the firm called both leads “narrow.”
  • Full ranked second on GDPval-AA v2.1. Artificial Analysis said the result is within the confidence intervals of Claude Opus 5.5 and Claude Sonnet 5.5.
  • Full’s biggest gain over Muse Spark 1.3 came on Terminal-Bench 4.0, from 33% to 57%, level with Gemini 4 Argon.
  • On the Artificial Analysis Coding Agents Index, Full scored 68 in Muse Code. That matches Claude Code with Sonnet 5.5 (68) and edges Opus 5.5 (66). Muse Wonder scored 63.
  • Knowledge “remains a relative weakness.” On AA-Omniscience, Full scored 30 and Muse Wonder 25, behind Opus 5.5, GPT-6 Astra and Gemini 4 Argon.
  • Hallucination rates are low. Full’s rate fell to 23%, second lowest among models scoring 45 or more on the index, behind Gemini 4 Argon. Muse Wonder was third at 27%.

Taken together, the outside numbers support the “frontier” label for Full. They are less clean on coding. On Artificial Analysis’s coding agent index, Full ties the best Anthropic setup rather than clearly beating it. And the model that builders can buy by the token, Muse Wonder, scores four index points and five coding-agent points below Full.

We did not review LMArena or other crowd-ranked leaderboards for this story.

Safety findings in the model card

Meta assessed Muse Wonder Full under its Meta Superintelligence Scaling Framework. The model card says that without mitigations, “we do not rule out the model reaching this threshold in the Chemical & Biological domain,” meaning the framework’s high-risk threshold. Meta says it applied safety post-training and system-level classifiers for that domain.

Meta rated the unmitigated model “moderate or lower” for cybersecurity and loss of control. With safeguards on, it says all domains reach “moderate or lower,” which it says supports deployment.

One detail stands out for security teams. On ExploitGym, Meta says Full captured 92 flags. A verification pass found 21 used the intended vulnerability and 71 used unintended ones. Meta notes that only 21 is comparable to the published ranking.

The model card also recommends extra guardrails, including “human-in-the-loop confirmation for irreversible actions,” when the model runs in agentic settings. It lists known limits: errors in multi-step reasoning, weaker visual and spatial reasoning (especially 3D), and uneven performance across languages. These are Meta’s own self-assessments, not outside audits.

What builders actually get

Here is the offer as it stands.

If you build on APIs, you get Muse Wonder. It has a 1M-token context, multimodal input and text output. It costs $4 in and $20 out per million tokens, with $0.20 cached input. Independent tests rank it near, but just below, the top tier, at a per-task cost above OpenAI’s and Google’s flagships and below Anthropic’s.

If you want Full, you buy into Muse Code. That means Meta’s tool, Meta’s plan limits and Meta’s five-hour and weekly resets. You cannot route Full through your own agent framework or a cloud marketplace.

If you run long agent jobs, the input bill is the number to watch. Artificial Analysis’s data suggests Muse Wonder reads a lot. Cache aggressively.

If you need audio, test it first. Meta’s own developer post tells you to confirm audio support on the release surface.

Meta’s frontier narrative

Meta frames this launch as a step toward “personal superintelligence.” Its research blog says: “Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence.”

Meta’s research blog says the models were “Trained at our gigawatt-scale Prometheus data center,” and: “Fourteen months ago, we began a ground-up overhaul of our artificial intelligence efforts. In that time, we’ve gone from rebuilding our stack from first principles to pushing the frontier.” TechCrunch, separately, reported this is the first frontier release since Scale AI founder Alexandr Wang restructured Meta’s AI efforts in August of last year, and that the models came from Meta’s Superintelligence Lab. The Ohio detail for Prometheus is TechCrunch’s reporting; Meta’s post does not name the state.

Reuters also reported that Meta shares rose about 4% in afternoon trading on launch day. It said the launch follows CEO Mark Zuckerberg’s pledge of billions of dollars and aggressive hiring of researchers over the past year.

The narrative is about personal agents and memory. The product, for now, is a coding subscription and a premium API model. Meta’s pitch for “active memory,” the ability to carry context across days and weeks, rests mainly on an internal benchmark. Until outside testers can run it, that claim is unverified.

Open questions

  • What do the Power and Maximum Muse Code plans cost? Meta has not published those prices in the posts we reviewed.
  • Will Muse Wonder Full come to the API? Meta says it is subscription-only. It has not said whether that will change.
  • When does Muse Wonder start powering Muse for subscribers? Meta’s research blog says “soon,” with no date.
  • Does audio input work on the API today? The model card lists it. The developer post says to confirm it.
  • Will outside labs reproduce the DeepSWE and SWE-Marathon leads? Meta’s rows mix sources and harnesses.
  • Can anyone else test active memory? MemoryBench is internal.
  • Will Muse Wonder be offered on cloud marketplaces? We found no announcement.

—

Sources

Cowry Tech News is published by Lone Cowry, an East Coast company.