Meta has released Muse Wonder and Muse Wonder Full, two new Artificial Intelligence (AI) models. They primarily target people who need to handle complex, long-running tasks, analyze large collections of material, or delegate coding work to an AI agent. The launch matters because this is Meta’s first frontier model, meaning a model intended to compete near the top of major performance rankings.
Meta’s two Wonder variants
Muse Wonder is the broadly accessible variant and can be used through the Meta Model API. An Application Programming Interface (API) connects a model to other software. Muse Wonder Full is more capable according to the published results, but the matching accounts from TechCrunch’s launch coverage and The Verge say it is currently restricted to Muse Code subscribers.
Both variants support a context window of one million tokens. A token is a short unit of text, while the context window determines how much material a model can consider within one task. The models can also accept text, images, video, and audio as input. In principle, that makes them suitable for work that combines meeting recordings, documents, and visual material.
Meta highlights two capabilities. “Active Memory” is designed to retain important information across conversations and retrieve it later. “Active Focus” describes the model’s ability to persist with difficult tasks, try alternative approaches, and ask follow-up questions. According to Meta, both features are intended for personal AI agents that can work with calendars, messages, and documents.
The models are also expected to power the Muse agent, initially for paying subscribers. Meta has not provided a firm release date or specific details about when specialized versions may reach a broader audience. The Verge similarly notes that Meta’s expansion plans currently lack a timeline.
Performance against leading rivals
Wonder represents a clear improvement over Muse Spark 1.3 in general performance testing. In the Artificial Analysis Intelligence Index v4.3, a combined measure covering several abilities, Muse Wonder Full scores 56 points and Muse Wonder scores 52. Muse Spark 1.3 scored 48. Benchmarks are standardized tests, but they never reproduce every condition you will encounter in daily use.
The biggest gains appear in agentic tasks, where a model must plan and execute several steps with limited supervision. According to The Decoder’s analysis, the reported Terminal-Bench 4.0 result rose from 33 to 57 percent. Muse Wonder Full placed among the top three models in six of the ten benchmarks examined.
The improvement is particularly strong in coding. Muse Wonder Full reaches 68 points in the Coding Agents Index, matching Claude Code with Sonnet 5.5. Muse Wonder scores 63, compared with 54 for Muse Spark 1.3. The lead report also says Wonder Full leads DeepSWE and SWE-Marathon, two evaluations of independently completed software tasks.
A concrete workplace example would be making changes to a larger software project. The model could analyze an assignment, inspect files, implement changes, and try a different approach after a failed attempt. That is substantially more demanding than generating one line of code in a chat window. The benchmark results suggest that Wonder Full is competitive on these longer workflows, but they do not establish reliability within a particular company or codebase.
Factual knowledge remains a distinct weakness. In the AA-Omniscience knowledge test, Wonder Full scores 30 points and Wonder 25, while Claude Opus 5.5 reaches 46, GPT-6 Astra 43, and Gemini 4 Argon 42. Wonder Full reportedly reduced its hallucination rate, the share of plausible-sounding but false responses, from its predecessor’s 33 percent to 23 percent. A lower error rate helps, but it does not entirely compensate for a smaller body of accessible knowledge.
Meta says the models achieve overall results comparable to Claude Fable 5.1 and GPT-6 Astra. This claim partly relies on Meta’s internal testing and should therefore be treated cautiously. The Verge also says it remains unclear how Wonder compares with Gemini 4 Argon. Based on Artificial Analysis scores, Simon Willison places the maximum Wonder variant around GPT-6.1 Sol, but his own image trials involving a pelican riding a bicycle are better viewed as an illustrative comparison than a rigorous quality test.
Pricing and access
Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API. Input tokens cover your instructions and submitted material, while output tokens cover the model’s generated responses. Long reports, extensive analysis, and agentic workflows with many intermediate steps can therefore cost considerably more than brief questions.
TechCrunch says Wonder is priced at roughly the same level as GPT-6 Sol and Opus 5.5. The Decoder reports that API use is almost three times as expensive as Muse Spark 1.3. The higher price mainly buys stronger performance on complex work that the model carries out with greater independence; for simple summaries or occasional drafts, the improvement may matter less.
The supplied sources do not give a separate price for Muse Wonder Full. It is unavailable through the API and instead comes through Muse Code subscriptions. The MarkTechPost launch summary confirms Wonder’s API prices and access to Wonder Full through Muse Code.
There is a minor inconsistency in the reporting. The Decoder’s summary says both Wonder variants are almost three times as expensive as their predecessor when used through the API. Its detailed availability information, along with the reports from TechCrunch and The Verge, says Wonder Full is only available through Muse Code. Without a published API option, there is no reliable API rate to assign to the Full model.
Pros and Cons of Muse Wonder
Pros:
- Agentic work – The largest reported improvements involve complex assignments made up of multiple independently executed steps.
- Long context – A one-million-token window allows the model to process substantial documents and extended work histories.
- Multiple media types – Text, images, video, and audio can be combined as input.
- Strong coding scores – Wonder Full reaches the level of leading competitors in the cited coding evaluations.
Cons:
- Weaker factual knowledge – Both variants trail several competing models by a wide margin in AA-Omniscience.
- Higher cost – API access to Wonder is almost three times as expensive as Muse Spark 1.3.
- Restricted access – Wonder Full is currently limited to Muse Code subscribers and has no stated standalone price.
- Limited independent testing – Several performance claims come from Meta and have not yet been comprehensively verified in everyday use.
Wonder in everyday use
For beginners
If you are trying Wonder for the first time, start with a clearly defined task whose result you can verify. For example, you could ask it to summarize a long document and then check whether its claims appear in the original material. This shows you the benefit of the large context window without immediately depending on long-term memory or a mostly autonomous agent.
A combination of media can provide another practical test. You could analyze meeting notes together with related images or a recording, provided your access method supports those inputs. The sources confirm technical support for these formats, but they offer no independent comparison of audio and video analysis quality.
For advanced users
If you already work with AI tools, the largest potential gain lies in multistep assignments. For a software project, you could have Wonder investigate an error, identify the affected files, propose a change, and check the result. The reported coding and agent benchmarks make Wonder, and especially Wonder Full, look well suited to this kind of workflow.
For longer personal workflows, Active Memory could reuse information from previous conversations, documents, or appointments. That may help with recurring project meetings because the agent is intended to combine past decisions with unresolved tasks over time. Persistent memory also increases the risk that confidential or outdated information will resurface in an unsuitable context.
The reports provide no Switzerland-specific availability details and no assessment of quality in German, French, Italian, or Romansh. They also do not explain where Active Memory data is stored, which privacy terms apply, or how users can control retained information. It therefore remains unclear how well the service fits internal data protection requirements when working with business documents, messages, or calendars in Switzerland.
Muse Wonder moves Meta visibly closer to the leading models for complex agentic and coding tasks. Its API price may be justified when the model completes long workflows reliably enough to replace several manual steps; for basic writing tasks, paying almost three times the Muse Spark rate is harder to defend. The main unresolved risk is the combination of still-limited independent testing, weaker factual knowledge, and persistent memory whose privacy framework is not explained in the supplied reports.
Sources
- Meta Launches Muse Wonder and Muse Wonder Full: Benchmarks, Pricing and How to Use It – MarkTechPost, 2026-10-13
- Metas neues Frontier-Modell Muse Wonder schafft den Sprung in die KI-Spitzengruppe – The Decoder, 2026-10-12
- Meta releases its first frontier model, Muse Wonder – TechCrunch, 2026-10-12
- Meta has some Wonder Full new AI models – The Verge, 2026-10-12
- llm-meta-ai 0.2 – Simon Willison’s Weblog, 2026-10-12
Image: Daniil Komov via Pexels