Add Muse Wonder to Your Routing Strategy

What are Muse Wonder and Muse Wonder Full?

Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026, as frontier models for personal and professional agents. They target work that involves planning, tools, remembering constraints, and making progress through multiple attempts. Both have a 1M-token context window. Meta positions Full as its most capable model; Wonder is the version developers can integrate into an API-based router. Meta’s announcement and model specifications describe the family.

For a routing strategy, treat Wonder as a candidate for demanding, sustained agent sessions. Start by comparing it against your existing frontier route on completed tasks, including failures and retries. A benchmark lead is a reason to evaluate a model, not evidence that it will win on your particular workload.

The latest on Meta Wonder Full access: Meta makes it available exclusively through eligible Muse Code subscriptions. There is no published per-token API price or confirmed general API route for Full as of writing this guide (Oct 14th). Details are likely coming shortly.

Why did Meta build Muse Wonder?

Meta describes Wonder as the next step toward personal superintelligence: agents that can help people accomplish substantial goals across work and everyday life. Its launch post places the release after a fourteen-month rebuild of its AI stack and the scaling work introduced with Muse Spark. Training used Meta’s Prometheus data center, with further scaling across pretraining and reinforcement learning. Research background.

The motivation is broader capability as well as stronger benchmark results. Personal agents encounter unfamiliar combinations of goals, tools, and constraints. Meta reports that Wonder Full generalizes better to tasks outside its targeted post-training distribution than the Muse Spark 1.3 trend would predict, attributing that improvement to stronger base models. That is Meta’s research finding; it does not establish a universal ranking for production traffic.

What work is Meta focusing on?

Meta emphasizes four connected areas: domain expertise, Active Memory, Active Focus, and coding. Professional evaluations cover legal document work, financial research, health questions, business workflows, and computer use. Coding work includes large repositories, changes across multiple files, debugging, and repeated test-and-revision cycles. Meta’s developer launch explains the intended workflows.

Personal demonstrations include planning a run around someone’s circumstances, personalized gifts, and a wedding seating chart with complex constraints. The shared requirement is coordinating information and actions over time. These examples suggest useful evaluation cases, but they are demonstrations rather than guarantees of reliable automation.

Wonder also supports multimodal workflows. Verify the modalities supported by your chosen endpoint: Meta explicitly says audio-dependent use cases require confirmation on the release surface. Model capability, gateway support, and the tools supplied by your application are separate requirements.

Active Memory: bringing relevant history into the current task

Active Memory is Meta’s term for retaining important context across conversations and retrieving earlier details when they become relevant. For example, a coding agent may need to recall an architectural decision made days earlier; a personal agent may need a preference that the user did not repeat in the latest request. Meta evaluates memory with MemoryBench. Active Memory description.

The routing implication is to consider the session and its history, not just the latest prompt. A short request such as “continue the migration” may depend on a large amount of prior work. Moving it to a cheaper model without a useful handoff can lose requirements and increase retries.

Implementation guidance: do not assume that calling Wonder through a gateway creates a persistent, cross-conversation memory database. Supply authorized history or retrieval tools through your application. Keep durable goals, decisions, sources, and permissions in storage you control, with user isolation and deletion rules. Before switching models, transfer the relevant state and confirm that the new route can use it.

Active Memory is also different from prompt caching. Memory determines which information matters; caching can reduce the cost of processing repeated input. A cache hit does not by itself make an agent remember a previous conversation.

Active Focus: sustaining progress when the task becomes difficult

Active Focus describes the model’s ability to continue pursuing a difficult goal: change approaches, recover from blockers, and adapt to new information. Meta also trains the models to request clarification and check before important actions. FocusBench evaluates focus, while SWE-Marathon examines sustained execution on long software projects. Active Focus description.

That makes Wonder worth testing when an agent repeatedly gets stuck, abandons a task early, or needs many human interventions. Examples include a migration that uncovers incompatible dependencies or a research workflow that needs to reconcile conflicting sources.

Persistence has a cost. Your agent runner should still enforce a session budget, tool-call limit, timeout, cancellation, and checkpoints for consequential actions. Measure whether the extra effort produces a verified result. More steps alone are not success.

Meta's Stance on Safety and Security

Meta says it trained the Wonder family to handle private information carefully and resist adversarial instructions. These protections matter when agents read messages, calendars, documents, and tool results that may contain instructions from someone other than the user. The launch announcement also describes its strongest safeguards to date against chemical, biological, radiological, nuclear, and cyber misuse, while retaining defensive cyber capabilities under its Superintelligence Scaling Framework. Meta’s safety statement.

These are model training and safety claims. The announcement does not establish immunity to prompt injection, a particular encryption design, zero data retention, or compliance certification. It says a fuller report with additional safety evaluations will follow. Keep those distinctions visible when assessing security features.

For routing, enforce data and tool permissions before selecting a model. Test hostile instructions embedded in retrieved content, leakage of private context, and whether the agent respects action approvals. Keep the same constraints on fallback routes. Model safeguards complement application controls; they cannot enforce access boundaries that your application never implements.

Also distinguish the ordinary Muse Spark route from the cheaper Contributor tier: OpenRouter states that Contributor prompts and outputs may be used to improve Meta’s products. It is not an interchangeable discount for sensitive traffic. Contributor tier description.

Which tasks should you route to Wonder?

The following is a proposed starting policy based on the announced capabilities, not a result from our own comparative benchmark. Use representative workloads to decide where Wonder earns its place.

How does pricing compare?

We reviewed pricing for Wonder at launch and are seeing the following in the early days of availability. These are all uncached input and output token prices, in USD per million tokens for the stated service. OpenAI figures use standard short-context pricing. Open-weight examples use named hosting endpoints on OpenRouter.

Wonder matches standard Opus 5.5 at $4/$20. It costs 3.2 times Spark 1.3’s input rate and about 4.7 times its output rate. Against GPT-6 Astra’s standard short-context $10/$50, Wonder is 60% cheaper per token. Against GPT-6 Sol or GPT-6.1 Sol at $2/$10, Wonder is twice the price. “GPT-6” is a family, so the variant matters.

Compare the cost of a completed task

For an illustrative request with 100,000 uncached input tokens and 10,000 billed output tokens, Wonder costs $0.60, standard Opus 5.5 $0.60, Spark 1.3 $0.1675, GPT-6 Astra $1.50, and Sol $0.30 at the rates above. This holds token counts constant to isolate price; models can use different numbers of tokens and attempts to finish the same work.

Wonder’s published cached-input rate is $0.20 per million tokens. If all 100,000 input tokens in that example qualify as cache reads, the token cost becomes $0.22. Preserve useful repeated context where supported, but verify cache eligibility and reported hits rather than budgeting for a guaranteed hit.

Include billed reasoning tokens, failed attempts, tool charges, gateway fees, cache writes where applicable, and human correction time. OpenAI publishes higher long-context rates and separate processing tiers; Opus also has cache-write and fast-mode prices. Self-hosted open weights add hardware, utilization, and operations costs. A single headline token rate cannot capture these differences.

Run Wonder against your incumbent on the same tasks and tool environment. Track verified completion, cost per accepted result, elapsed time including p95, intervention count, and constraint violations. Promote it only for task categories where the measured quality or completion gains justify the spend.

Where can you implement it now?

You can start through OpenRouter or Vercel AI Gateway, as well as directly through Meta Model API. Both gateways currently list Meta as the upstream provider for Wonder. Two gateway integrations therefore do not necessarily provide independent model-hosting capacity. OpenRouter listing; Vercel listing.

What should you monitor next?

Watch for additional hosting providers, Wonder Full API access, published Full token pricing, and the additional safety evaluations Meta said would follow. Treat “coming soon” as a roadmap claim until a provider publishes an endpoint, price, and usable access terms. No unverified provider launch dates are assumed in this guide.

For each new listing, record the exact model version, upstream host, supported modalities and tools, region, retention and training policies, caching rates, and account restrictions. Then run a small compatibility and quality evaluation before changing routing policy. Update this guide when those facts change; a catalog entry alone is not evidence of a tested production route.