Muse Wonder explained

Muse Wonder is Meta's new model family for coding and long-running AI agents, announced on October 12, 2026. The practical choice is between Muse Wonder, available through the Meta Model API in preview, and Muse Wonder Full, the higher-capability version offered through eligible Muse Code subscriptions.

For an application integration, start with Wonder's documented model ID, muse-wonder-1.0. Its direct API price is $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Full has no public pay-as-you-go API offering at this check date. Setting Wonder's reasoning effort to max does not switch to Full.

Muse Wonder vs Full: which can you use?

These are Meta's published access terms, not a guarantee of regional or account eligibility. The model page lists the subscription prices; check your checkout for available benefits and final pricing. Plans have usage limits.

Three names describe different layers. Muse Wonder is a model; Muse Code is a terminal coding agent; Muse is the personal-agent app. Meta's announcement says Wonder will roll out to the personal app, subscribers first. API availability does not establish that your app account already uses it. Our Muse Agent review covers the app, invitations and privacy controls.

What changed from Muse Spark?

Meta positions Wonder for longer, harder tasks: changes across a repository, tool use, computer operation and recovery after failed attempts. The API accepts text, images, video, audio and PDFs and produces text. Image or video understanding does not make this an image or video generator.

Two terms anchor Meta's research announcement:

• Active Memory: retaining important context across conversations and retrieving relevant details.
• Active Focus: continuing toward a goal, trying alternatives, recovering from blockers and checking with the user before important actions.

These describe trained capabilities. A single API request does not create a complete personal agent with connected accounts and durable memory. Your application still needs to supply context, manage state and execute permitted tools. The 1M context window measures how much a request can hold, not how long your application remembers something.

Meta recommends Wonder for demanding agent work and Spark for volume, latency or cost-sensitive tasks. That is a useful evaluation split; it is not a measured speed comparison from us.

Benchmarks: better than Spark, with a higher task cost

Artificial Analysis's October 12 evaluation provides an independent comparison. All three columns below use max reasoning effort; the Intelligence Index is v4.3.

The evidence supports a meaningful improvement in agentic work over Spark. It also shows why a token price alone is insufficient: Wonder's estimated index-task cost is about 2.8 times Spark's in this evaluation. Full's stronger result cannot be assigned to the smaller API model.

The $4.52 figure is an evaluation-specific estimate, not the price of a coding request or your production task. Model, reasoning level, harness, cache use, retries and task mix all affect the result. We have not measured Wonder's latency or success rate on an OmniaKey workload.

API pricing: calculate cached input and reasoning correctly

The direct Meta rate card lists these USD prices per 1 million tokens in Standard mode:

Meta lists no long-context premium. Web search grounding adds $2.50 per 1,000 search queries, beyond token charges. Wonder has no Contributor-mode API variant at this check date.

For the Responses API, read total input from usage.input_tokens and cache hits from usage.input_tokens_details.cached_tokens. Cached tokens are already part of total input. Subtract them before applying the uncached rate. Reasoning tokens count toward billed output, even though they are not visible answer text.

USD = ((total_input - cached_input) × 4
       + cached_input × 0.20
       + billable_output × 20) / 1,000,000

For 100,000 total input tokens and 10,000 billable output tokens, the token charge is $0.60 with no cache hit. If 90,000 of those input tokens hit the cache, it becomes $0.258. Ten web-search queries would add $0.025 to either scenario.

These are worked assumptions, not measured jobs. Cache hits are automatic when prefixes match but are not guaranteed. An agent making multiple calls must sum every call, including retries and reasoning. Compare cost per accepted result, not just the cost of its final answer.

How to call the Muse Wonder API

Create a Meta Model API credential in the official developer console and make it available locally as MODEL_API_KEY. The official overview documents the following Responses contract:

curl --fail-with-body --silent --show-error \
  https://api.meta.ai/v1/responses \
  -H "Authorization: Bearer $MODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "muse-wonder-1.0",
    "reasoning": {"effort": "high"},
    "max_output_tokens": 4096,
    "input": "Outline a SQLAlchemy 1.4 to 2.0 migration. List breaking changes, a safe order of work, and tests to run."
  }'

This is a direct Meta request, requiring a Meta credential. The output limit includes reasoning; inspect response status and usage before assuming the visible answer is complete. Meta also documents Chat Completions and Messages surfaces, but their fields differ: Responses uses reasoning.effort, while Chat Completions uses reasoning_effort.

If a request fails, check the base URL, credential, exact model ID and your account's access. Do not guess a Full API alias or append -contributor to Wonder. A Muse Code subscription does not convert an ordinary API key into prepaid or unlimited API access.

Is Muse Code Full worth a subscription?

Full is worth evaluating if you want Meta's highest-capability coding model and can work inside Muse Code. The listed Power Usage plan is $50/month; Maximum Usage is $200/month. Those are subscription prices, not an API token allowance you can transfer to another client.

The subscription documentation says the plan applies to the Muse Code credential connected during onboarding. Additional API keys remain pay-as-you-go. Subscriptions default to Contributor mode, which allows content to be used to train future Meta models in exchange for more usage; it can be turned off in settings. Wonder's Standard API mode says content is not used for training.

Before buying or migrating, repeat one real repository task against your existing setup: use the same starting revision, instructions, tools and acceptance tests. Record completion, review time, retries, elapsed time and billed usage. This tests the benefit you are buying more directly than a leaderboard position.

Common questions

Is Muse Wonder open source or available to download?

The official material checked here offers hosted API and Muse Code access, without downloadable Wonder weights or a self-hosting license. Do not infer an open-weight release from Meta's Llama history; Meta documents Muse Glimmer separately as an open-weight model.

Can I use Muse Wonder for free?

The published API is pay-as-you-go, and Full requires an eligible paid Muse Code plan. We did not verify a universal free allowance. A free Muse app download or an older referral reward does not establish free Wonder API access.

Does max reasoning unlock Wonder Full?

No. Meta explicitly distinguishes reasoning effort from the model variant. muse-wonder-1.0 at max remains Muse Wonder.

Can I use Muse Wonder through OmniaKey?

We did not find a Wonder entry in the OmniaKey catalog checked on October 14. Use the current model catalog and pricing page to verify supported routes. The Meta request above is not an OmniaKey endpoint or price quote.

Which version should I choose first?

Choose Wonder's official API for an app integration, evaluate Full through Muse Code for demanding repository work, and retain Spark as a lower-cost baseline for simpler or high-volume tasks. The decision should follow your completed-task results and actual access.

Sources

• Meta launch announcement, October 12
• Meta models and availability
• Meta model page and subscription list prices
• Meta API prices and rate limits
• Artificial Analysis independent launch evaluation

Sources are in English. Access and prices were checked on October 14, 2026; confirm the current documentation and account terms before purchase.