Muse Wonder is worth evaluating when your application needs a model for substantial coding or multi-step agent work. The first integration decision is access, not a benchmark chart: muse-wonder-1.0 is on Meta Model API in Preview; Wonder Full is not a public pay-as-you-go endpoint. A Muse Code subscription and an application API key are different purchases.
This is a documentation-based integration guide, checked October 13, 2026. We have not called the paid endpoint, measured latency or reproduced Meta’s benchmarks. It explains how to prepare a controlled evaluation, not that this Preview model is already proven for your production workload. Use APIpulse’s Pricing Index for maintained cost data and Kingy’s Wonder-versus-Full analysis for the broader variant and benchmark buying decision.
Choose the right access path
| Need | Path | Boundary |
|---|---|---|
| Call a model from your own backend | Meta Model API, muse-wonder-1.0 | Preview, token billing |
| Use Meta’s terminal coding product | Muse Code | Product permissions and plan limits matter |
| Access Wonder Full | Eligible Power Usage or Maximum Usage Muse Code subscription | No public token-priced Full endpoint |
| Serve downloaded weights locally | Not Wonder or Full | Evaluate a separately licensed open-weight model |
The Meta model documentation lists a 1,048,576-token context for regular Wonder. Inputs include text, images, video, audio and PDFs; output is text. Input understanding does not mean this endpoint generates images or synthesizes audio. It supports tools and structured output, but your application still validates and authorizes actions.
Selecting a higher reasoning effort on Wonder does not turn it into Full. Do not guess a full model ID or assume an account subscription credential grants arbitrary API use. For terminal installation and workflow questions, keep the separate Muse Code owner as your starting point.
Prepare a small direct API request
Meta’s overview gives the base URL https://api.meta.ai/v1 and bearer authentication using a model API key. First verify your account’s model access, billing and geographic eligibility in Meta’s developer interface. Public documentation is not proof that a particular account is enabled.
Keep the key on the server, outside source control and browser bundles. Follow our environment variables foundation for local, preview and production separation. Start with a non-sensitive prompt and a small output cap, not a repository upload or a tool that can write files.
The following is a documentation-derived request shape, not an executed test:
curl --fail-with-body https://api.meta.ai/v1/chat/completions \
-H "Authorization: Bearer $MODEL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "muse-wonder-1.0",
"messages": [{"role": "user", "content": "Explain an idempotent job queue in three sentences."}],
"max_completion_tokens": 256
}'Compare the current quickstart and Chat Completions protocol before running it. Meta exposes Chat Completions, Responses and Messages compatibility; compatibility does not guarantee every extension in an existing SDK or agent harness works unchanged. Begin with a simple request, then test the exact streaming, tool and schema features your application uses.
Chat Completions does not preserve conversation state between requests: your application supplies the relevant message history. For multi-turn reasoning workflows, evaluate the Responses protocol and its previous_response_id mechanism instead of assuming a Chat request inherits earlier reasoning. Keep application-owned task state separate from either protocol.
Keep a thin provider adapter. It should translate authentication, request fields, returned usage and errors into your application contract. Do not spread model-specific assumptions across every workflow. Our AI API design foundation explains how streaming, jobs and tool responses fit that boundary.
Budget for the actual workload
Published Standard-mode rates are $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Meta lists no long-context premium. Search grounding is separate at $2.50 per 1,000 queries, plus token costs. Wonder has no Contributor-mode API price.
As an illustrative calculation, 100,000 uncached input tokens and 20,000 billable output tokens cost $0.80, before search, retries and other applicable charges. That is arithmetic, not a measured coding-task cost. An agent may call repeatedly, reread changing files and generate substantial intermediate work.
Record uncached input, cache hits and billed output separately. A quoted cache rate is useful only for tokens actually served from cache. Check reported usage rather than assigning every repeated prompt a discount. A large context allowance is not a recommendation to submit the whole repository on every turn.
Use the cross-provider pricing comparison to shortlist alternatives, then measure cost per accepted task. Avoid a pairwise “cheaper” verdict based only on output rates: failed runs, tool retries and review time can dominate the total.
Rate limits are shared across the team
Meta lists 3,000 requests and 4 million tokens per minute for Wonder, per team, not per API key. A developer’s smoke test and a production worker can consume the same allowance if they share a team. Do not solve throttling by multiplying keys and assuming each gets a fresh quota.
Use bounded queues, concurrency limits and backoff with jitter. Honor a returned retry delay. Separate a retriable transport problem from authentication, malformed input or unavailable-model failures. The AI API failure guide covers safe responses and fallbacks.
For a tool workflow, a retry also needs an action policy. Replaying an entire job after a model timeout can send the same email or edit the same record twice. Keep idempotency and human approval around consequential operations, even when the model appears capable of completing the sequence itself.
Active Memory is not your database
Meta describes Active Memory as carrying useful context and Active Focus as maintaining progress through difficult work. Treat these as model capabilities to evaluate, not an unlimited durable-state guarantee. The launch’s performance claims remain attributed to Meta and its named benchmark sources; this guide makes no independent ranking claim.
Your application needs explicit state ownership. Store task IDs, approved decisions, tool results and completion status in your own system. Test what happens after a restart, a contradictory instruction or a deleted conversation. A successful answer with a long prompt does not establish perfect recall across unrelated future sessions.
Keep the agent’s permissions narrower than its language capabilities. Start with read-only tools, then introduce one reversible action behind a review gate. Validate returned arguments and structured responses using the testing and evaluation foundation. Never let a generated shell command inherit unrestricted production access by default.
Privacy checks before using real code
Meta says Standard-mode content is not used for model training. Wonder’s direct API is Standard-only. Muse Code subscriptions instead default to Contributor mode, which permits content use for training and can be switched off. Verify the actual product and setting; do not transfer a privacy assumption from one access path to another.
Training exclusion is not proof of zero retention, local processing or your required data residency. Review terms, account settings and any contractual controls before uploading proprietary code or customer documents. Use redacted fixtures for the first evaluation. Our AI security hub connects authentication, secrets and permission boundaries.
A production evaluation that answers a decision
Build a small representative dataset before choosing a model. Include ordinary tasks, malformed inputs, ambiguous requirements, tool failures and cases where the correct action is to ask for approval. Score correct completion, not merely plausible prose. Record the exact model, settings, prompts and tool definitions so a later rerun means something.
For coding, run generated changes in an isolated branch or sandbox. Check compilation, tests, security-sensitive diffs and maintainability. For a document workflow, verify cited evidence and extracted values. Measure task completion time, error rate and total charges independently; a model can improve one while worsening another.
Use the LLM regression testing guide before a provider/model change and the AI Operations hub for monitoring and rollback. Keep a working fallback, a budget ceiling and a way to stop an agent. Preview is an additional reason to avoid an irreversible all-traffic migration.
My take
Evaluate Wonder for demanding API-driven agent work if you can tolerate Preview change and have a task-specific quality gate. Do not buy Full expecting a drop-in backend endpoint, and do not choose either variant solely from a launch leaderboard. Cheap, repetitive sub-tasks may fit a lower-cost model; sensitive offline work may require open weights instead. The right next step is a bounded evaluation, not a universal migration recommendation.
FAQ
Is Muse Wonder Full available through the API?
Not through a public pay-as-you-go Meta endpoint in the reviewed documentation. Eligible Muse Code subscriptions provide Full access.
Does the million-token context increase the input rate?
Meta lists no long-context premium for Wonder. More submitted tokens still increase the bill, even without a higher unit rate.
Is Muse Wonder downloadable for local inference?
No public Wonder or Full weights are available in the reviewed release. Muse Glimmer is a separate open-weight family, not downloadable Wonder.
Does identity or subscription access change API privacy?
Check the access mode rather than inferring privacy from the account. Direct Wonder API Standard mode and subscription Contributor mode have different training treatment.
Can I ship without a hands-on model test?
Documentation can establish integration facts, not production reliability. Run your own evaluation and review before introducing sensitive data or consequential tools.