Meta Muse Wonder: 1M-Token Agent Model Arrives via API

Introduction

On October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, which Meta's research blog presents as "A Frontier Model For Personal & Professional AI Agents." The model card gives October 12, 2026 as the release date and lists the provider as Meta Platforms. Muse Wonder is available through the Meta Model API at dev.meta.ai and through Muse Code. Muse Wonder Full is limited to Muse Code subscribers and, according to TechCrunch, is not served through the API. Meta also says Wonder will soon power the Muse agent, with subscribers getting it first, and that it was trained at its gigawatt-scale Prometheus data center, which TechCrunch places in central Ohio.

The release matters because it is Meta's pitch for a general agent model rather than a chat model. Earlier Evermx coverage looked at the Muse Spark launch, Muse Code and the Muse Gadget SDK; this article focuses only on the Wonder model family and what Meta's own model card says about it.

Feature Overview

1. Specifications from the model card

2. Agent-first intended uses. The card says the models are intended for commercial and research use, with particular strengths in personal and professional agentic tasks. Its list of use cases covers general assistance, coding and software engineering, personal and professional agents using tools and applications, domain knowledge work in law, finance and health, multimodal understanding of documents, screenshots, charts and video, and scientific research support.

3. Active Memory and Active Focus. Meta's announcement names these two concepts as the design themes of the model. Active Memory means carrying important context about the user across conversations and retrieving the details relevant to the task at hand. Active Focus describes how the model keeps working toward a hard goal: trying alternative approaches, working through blockers, asking for clarification when needed and checking in before important actions. The model card describes the evaluations behind each theme. MemoryBench is an internal test of recalling critical facts across multi-turn conversations of 64K to 100K tokens that include distractor conversations and injected turns that conflict with or later update the target details. FocusBench is an internal evaluation of personal agentic tasks in areas such as housing, travel, IT and finance, measuring whether an agent persists through real-world blockers across multiple interactions.

4. A broad but mostly internal benchmark suite. The card compares Wonder with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 at max reasoning effort. Named benchmarks include the Artificial Analysis Intelligence Index v4.3, the Personal Superintelligence Index (PSI), AutomationBench-AA, Harvey LAB-AA, Vals Finance Agent V2, HealthBench Pro, Agents' Last Exam, OSWorld V2.1, Terminal-Bench 4.0, DeepSWE v1.1, SWE-Marathon and ProgramBench. PSI is described as an internal composite of seven evaluations: Agents' Last Exam, AutomationBench-AA, FocusBench, JobBench, GDP.pdf, VBT and DRACO. Meta says the series reaches state-of-the-art results on personal and professional agent, memory and coding benchmarks. TechCrunch reports that Wonder and Wonder Full score comparably to Claude Fable 5.1 and GPT-6 Astra, and that Wonder beats Fable, Opus and Astra on MemoryBench in Meta's testing. The capability scores appear in the blog's charts rather than in the model card text, and MemoryBench, FocusBench and PSI are Meta-internal benchmarks, not public academic ones.

5. Published risk evaluations for the Full variant. The card reports risk results for Wonder Full, its most capable variant, under the Meta Superintelligence Scaling Framework. Cyber capability is measured with CyberGym, ExploitBench and ExploitGym, adversarial robustness with CyberBench, PolyGuard and RedCode, chemical and biological risk with SecureBio and LAB-Bench tests plus BioTIER refusal coverage, and loss of control with MASK and Agentic Misalignment. Meta says Full demonstrates strong robustness to adversarial attacks across its cyber benchmarks and strong refusal performance in English, across 10 language buckets and under adversarial prompts. Among the reported figures, Wonder Full shows a 0% harmful-action rate on Agentic Misalignment, 87.3% on the MASK honesty test, and BioTIER refusal rates of 98.7% in English and multilingual testing and 98.2% under adversarial prompts.

Usability Analysis

For developers the practical entry point is the Meta Model API, where Wonder is priced at $4 per million input tokens and $20 per million output tokens according to TechCrunch. The 1M-token context and four input modalities make it suitable for long documents, screenshots and recorded video, although output is text only. Muse Code users get the same family inside a coding product, and subscribers can reach Wonder Full there.

The card's methodology notes matter for anyone comparing numbers. Meta says it uses a comparable third-party or provider-reported result where one exists, and otherwise takes the highest of its internally validated run, a self-reported score or an official leaderboard score. Competitor runs are described as best-effort, with prompts, tools and runtime possibly not tuned for proprietary models. For SWE-Marathon, the leaderboard results for GPT-6 Astra and Claude Fable 5.1 use Codex and Claude Code respectively, so Meta says that row is not a controlled common-harness comparison. Buyers should rerun their own workloads before treating any cross-vendor ranking as settled.

The ExploitGym note is a good example of candid reporting: Meta says a verification pass found most captured flags came through unintended vulnerabilities, and that only the intended-vulnerability count is comparable to the published ranking.

Pros and Cons

Pros: a 1M-token context with text, image, video and audio inputs; a detailed model card with per-benchmark methodology; disclosure of caveats such as the non-controlled SWE-Marathon row and best-effort competitor runs; a Wonder tier reachable through a public API.

Cons: the strongest variant is gated to Muse Code subscribers; a full safety report is still pending; several headline benchmarks are Meta-internal, and Meta describes competitor runs as best-effort; Meta lists visual and spatial reasoning, especially 3D tasks, as a known limitation.

Outlook

Meta says Wonder will soon power the Muse agent, with subscribers first, which would move the model from developer tooling into a consumer assistant. The pending safety report is the next document to watch, because the card says that without mitigations Meta does not rule out the model reaching its high-risk threshold in the Chemical and Biological domain. With safety post-training and system-level classifiers, Meta assesses residual risk at "moderate or lower" across its three catastrophic-risk domains, and says an internal governance process reviewed that assessment before deployment. Independent replication of the agent benchmarks will show how much of the reported lead holds outside Meta's harnesses.

Conclusion

Muse Wonder is a credible entry for teams building long-context, tool-using agents, and its model card documents per-benchmark methodology and caveats in detail. Its value depends on independent verification of benchmark claims, on the forthcoming safety report, and on whether Wonder Full ever reaches the API. Developers and agent builders should pilot the Wonder tier; buyers seeking the top variant must currently subscribe to Muse Code. Rating: 4 out of 5.

Editor's Verdict

Meta Muse Wonder: 1M-Token Agent Model Arrives via API earns a solid recommendation within the Other LLM space.

The strongest case for paying attention: 1M-token context with text, image, video and audio inputs. That alone raises the bar for what readers should expect in this space. Reinforcing that, detailed model card with per-benchmark methodology and sourcing — practical value rather than just headline appeal. The broader signal worth registering is straightforward: a model card that explains its score-selection rules lets readers judge how much weight each benchmark deserves. On the other side of the ledger, one constraint is real rather than a marketing footnote: the strongest variant, Wonder Full, is limited to Muse Code subscribers and is not served through the API according to TechCrunch. It should factor into any serious decision. Layered on top of that, a full safety report is still pending, which Meta says will be available soon, so the complete risk picture is not yet public — which narrows the set of teams for whom this is an obvious yes.

For agent builders, teams running long-context or multimodal workloads, and anyone already comparing frontier models through an API, the Wonder tier is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to wait for the full safety report and independent benchmark replication.