Evidence at a glance
More Than a Larger Model Tier
Meta Superintelligence Labs launched Muse Wonder and Muse Wonder Full on October 12, 2026, positioning them for personal and professional agent tasks. Both have a one-million-token context window. Wonder is available through the Meta Model API and Muse Code, while Wonder Full is restricted to Muse Code subscribers. Meta says the work took roughly 14 months from rebuilding its AI efforts to this release.
The launch matters not because a new model name proves Meta has joined the frontier, but because it targets a weakness the company had previously acknowledged: long-horizon agents and coding workflows. After Muse Spark, Meta is shifting the product question from what a model can answer to whether it can keep working until a task is done. That is a change in direction; whether it becomes a durable advantage depends on completion rates, cost, and reliability in real work.
The Gains Cluster Around Multi-Step Work
Meta’s description focuses not on a newly disclosed architecture, but on how the models keep work moving: planning multi-step tasks, using tools in parallel, delegating subtasks to agents, and changing course after errors or blockers. Meta calls retaining and retrieving relevant context across conversations Active Memory. It calls adapting to new information, trying another route, or asking for clarification when stuck Active Focus. Public materials do not disclose a specific training recipe, parameter count, or reproducible implementation details, so these labels are best read as product descriptions rather than independently verifiable accounts of internal mechanisms.
Third-party evaluations offer a more concrete signal. On Artificial Analysis’s Intelligence Index, Wonder Full scored 56, Wonder 52, and Muse Spark 1.3 48, while Opus 5.5 scored 58. Wonder Full rose from 33% to 57% on Terminal-Bench 4.0 and from 1681 to 1857 Elo on GDPval-AA. The gains are concentrated in tests involving tools, multiple steps, and task execution, making them more informative than a general claim that the models are simply smarter.
The Harness Changes the Result
Coding evaluations show the limits of that progress. In Artificial Analysis’s Coding Agents Index, Muse Wonder Full paired with Muse Code scored 68, matching Claude Code with Sonnet 5.5 and exceeding Opus 5.5 at 66; Wonder scored 63. This is not a general ranking of standalone models, but a result for a model paired with a particular agent product. Meta’s model page also reports strong results on DeepSWE 1.1 and SWE-Marathon, while its model card cautions that reasoning settings differ and not every result comes from a fully controlled comparison.
That is why “Claude 5.5-tier” should be read as shorthand for competitive positioning, not certification of parity across the board. Wonder Full remains below Opus 5.5 on the overall index, yet matches Claude Code with Sonnet 5.5 on the coding-agent index. For technical leads, the unit of evaluation should therefore be more than the model name: it includes reasoning tier, tools, execution environment, and task. Change the harness, and both the ranking and practical value can change.
Product Tiers Also Set a Boundary
The split between Wonder and Wonder Full turns a capability distinction into an access distinction. Wonder is available through an API at $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens. Wonder Full has no published per-token price and is not available through the API; it is tied to Muse Code subscriptions. Power Usage costs $50 a month and Maximum Usage costs $200. For teams, the most accessible tier is not necessarily the one with the fullest reasoning capability.
That means “stronger” does not automatically mean “better to deploy.” Long context and multi-step tool use may help complete tasks, but can also increase the amount of input, calls, and execution involved. The available materials are not enough to estimate costs for every organization. Teams should measure total consumption on real workflows and check subscription limits, quota rules, and data-access arrangements before adding the model to production, rather than relying on benchmark scores alone.
A Quota Bug Is a Product-Maturity Signal
On launch day, users reported that Wonder and Wonder Full were consuming Muse Code quotas abnormally quickly. Meta fixed the issue that day and reset quotas for all users, while the team said other problems remained to be addressed. A single incident does not establish that the models themselves are unreliable. But it shows that once intensive reasoning is packaged as a subscription product, metering, quotas, and user experience become part of the system. For agents, the number of steps taken, recovery behavior, and visibility into consumption all affect whether users can trust the product to keep working.
The sound conclusion for now is that Muse Wonder offers competitive evidence in long-horizon tasks and coding agents, but does not prove across-the-board parity under every task, configuration, and cost condition. Teams considering a trial should start with reviewable tasks and constrained permissions, while tracking completion rates, human interventions, and actual token use. For email, files, or external actions, keep explicit authorization and confirmation points. Scale deployment only after task performance, cost, and quota behavior have been tested under real workloads.