Evidence at a glance
The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
One Model Family, Two Ways to Get It
Meta has introduced Muse Wonder and Muse Wonder Full as models for personal and professional agents, spanning coding, legal and finance work, and operation of browsers, desktop software, and mobile apps. Both are listed with a one-million-token context window. They accept text, images, video, and audio, return text, and run in Meta's cloud through the Meta Model API or Muse Code.
The first distinction is access, not a simple high-versus-low setting. Muse Wonder is available through the API and Muse Code, while Wonder Full is restricted to paid Muse Code plans. Meta's documentation also says that setting Wonder's reasoning effort to max does not turn it into Full. For engineering leads, evaluation therefore includes not only model capability but also cloud dependence, subscription access, and two different integration paths.
Memory and Focus Are Promises, Not a Disclosed Architecture
Meta puts two named behaviors at the center of the release. Active Memory is described as carrying user context across conversations and retrieving relevant details when a task calls for them. Active Focus means trying alternate approaches when blocked, continuing to work, and asking questions or seeking confirmation before important actions. The model card lists a context length of 1,048,576 tokens, which provides room to process large amounts of material in one run, but does not by itself explain how cross-session memory is stored or retrieved.
Meta also says the models can break down long tasks, delegate to sub-agents, call tools in parallel, and recover from errors. There is an important distinction behind this product language: these are descriptions of intended behavior, not a disclosed account of the memory or orchestration architecture, and the material does not show that every task will complete reliably. For engineering teams, “keeps trying” and “finishes correctly within the right boundaries” are separate acceptance criteria.
The Benchmarks Show a Profile, Not a Clean Sweep
Meta's results are best read as a capability profile. Muse Wonder Full scores 79.0 on DeepSWE 1.1, just above Muse Wonder's 78.8. On SWE-Marathon, Full scores 57.5 and Wonder 53.1. But Full's SWE-Marathon run uses xhigh reasoning effort, while rival results were produced with tools such as Codex or Claude Code. Those numbers are not a direct contest under a common setup.
The weaknesses are just as visible: on ProgramBench, Wonder scores 50.0 and Full 59.0, below Claude Opus 5.5 at 70.5 in Meta's table. On MemoryBench, Wonder scores 85.6 and Full 85.5, but that is an internal Meta evaluation. The cautious reading is neither “Full is simply better” nor “the models are good at coding.” Results vary by task and test conditions, so teams should distinguish third-party benchmarks, internal evaluations, and runs made with different tool configurations.
Price and Access Shape Where It Fits
Muse Wonder's API pricing is listed per million tokens: $4 for input, $0.20 for cached input, and $20 for output. Wonder Full has no published per-token price; it is included in Muse Code's Power and Maximum plans, listed at $50 and $200 per month respectively. Long tasks can accumulate both substantial input and substantial output, so the input rate alone cannot represent the cost of an agent run.
The API offers a direct programmatic route. The documented model name is muse-wonder-1.0, and Meta says the API is compatible with the OpenAI SDK. Muse Code places the models inside a subscription product, with Full access tied to plan eligibility. Teams that need controlled toolchains, predictable calling patterns, and traceable spend should assess the API first. A subscription may fit developers who want Full inside a coding environment, but quotas and task-level costs still need checking; a monthly fee alone does not tell you how much work it will support.
Treat the Long Context as Capacity, Not a Reliability Guarantee
The key question to test with Muse Wonder is whether it can connect context, memory, and tool use into a controlled task loop—not whether it tops one leaderboard or appears to keep working in a demonstration. A practical pilot should use familiar tasks with verifiable outcomes and record completion rates, human handoffs, whether the model seeks confirmation before consequential actions, and input and output costs. In high-impact settings such as legal work, finance, or production code, teams should also test permission boundaries and error recovery rather than treating a promise to ask before important actions as a governance mechanism.
A million-token window means the model can receive a large amount of material. It does not mean the model will always retain the right details, nor does it make cross-session memory accurate by default. Because the models run only in Meta's cloud, data handling, tool permissions, and vendor dependence also belong in the deployment decision. If a team cannot define which actions require human approval, how outputs will be checked, and how a task can be stopped when it fails, a longer context window or more persistent agent is not a reason to grant broader authority.