Meta Muse Wonder vs Full: Pricing, Access and Agent Limits

On October 12, 2026, Meta launched two frontier models for long-running AI agents: Muse Wonder and Muse Wonder Full. Both emphasize coding, memory, computer use and multi-step work, but they are not sold in the same way. Muse Wonder is available through Meta's Model API and Muse Code; the more capable Wonder Full is available through qualifying Muse Code subscriptions, not the pay-as-you-go API. For builders, that difference matters more than a leaderboard headline.

Muse Wonder vs Wonder Full: what is actually available?

Meta's developer page lists model ID muse-wonder-1.0 for API use. Wonder Full is listed as muse-wonder-full-1.0, but its availability is through Muse Code subscriptions rather than a public token-metered API. Meta says Wonder Full is its most advanced model; that does not mean an API customer can simply select it in production. The prices above are published list prices, not a guaranteed total bill, and may vary with plan terms and additional tools.

What a real API workload could cost

At the announced rates, a workload with 1 million uncached input tokens and 200,000 output tokens costs $8 in model tokens: $4 for input plus $4 for output. A larger workload with 10 million uncached input tokens and 2 million output tokens costs $80: $40 input plus $40 output. If every one of those 10 million input tokens qualified for the published cached-input rate, that example would be $42 instead. Actual cache eligibility, repeated prompts, tool usage and retries can materially change the total. Measure the complete agent run rather than estimating from a single prompt.

Active Memory and Active Focus explained

Meta describes Active Memory as carrying relevant context across conversations and retrieving it when useful. Active Focus is the model's ability to continue a difficult task, try alternatives, recover from errors and ask for clarification or approval before important actions. These are product capabilities described by Meta, not proof that an agent will remember everything correctly or complete every workflow unattended. For teams working with customer records, private documents or connected apps, memory and autonomy create new evaluation requirements around consent, retention and access controls.

What Meta's benchmarks do—and do not—show

Meta reports strong results across agentic computer use, long-horizon coding, legal, finance and memory evaluations, including a composite Personal Superintelligence Index. The benchmark mix includes Meta-defined measurements and external tests, and results depend on evaluation settings and reasoning budgets. The company's research blog also notes that unfamiliar tasks are an important test of generalization. Treat vendor comparisons as a reason to pilot, not as a universal ranking for your own CRM, repository or support workflow.

A practical five-step evaluation for builders

1. Choose 20–30 representative tasks: a multi-file code fix, document comparison, browser workflow and at least one task that should be refused or escalated.
2. Test Muse Wonder against your current model using the same tools, prompts, permissions, retrieval data and time limits. Do not compare the API model with Wonder Full as though both were API-accessible.
3. Record successful completions, human corrections, unsafe actions, latency, tokens, cache hit rates and tool costs. Calculate cost per correctly completed task.
4. For memory-enabled agents, test incorrect recalled facts, stale preferences, sensitive-data exposure and explicit revocation. Require approval for purchases, outbound messages or irreversible edits.
5. Start with a small permission-scoped pilot. Keep audit logs, a human override and a rollback route; expand only when real task success and governance meet your thresholds.

What should teams choose?

Muse Wonder is the relevant option if you need a callable model inside your own product or automation. Muse Wonder Full may suit developers who want a more capable coding experience within the Muse Code subscription, but it should not be included in an API architecture until Meta actually offers that access. Neither a one-million-token context window nor an agent's ability to operate apps removes the need for good retrieval, scoped credentials and testing.