Muse Wonder and Full: access, coding workflows and costs

Meta released Muse Wonder and Muse Wonder Full on October 12, 2026. Both are designed for personal and professional agents, from document work to coding projects that take many tool calls. The ordinary Wonder model is available through Meta Model API and Muse Code; Full is the higher-capability variant available through eligible Muse Code subscriptions.

Meta calls remembering relevant context across conversations Active Memory. Active Focus describes continuing toward a goal through blockers, alternative approaches and new information. In a coding workflow, that means keeping requirements in view across edits, tests and repairs. Your application still supplies the memory records, files and tools the agent can access.

The model card lists a 1M-token inference context and evaluates Wonder and Full separately. Meta’s demos include a playable kitchen game built from images and a browser-based running plan. Those are useful task examples; choose the same model variant and agent setup when trying the workflow yourself.

Create a Meta Model API account and an API key for your application. Point an OpenAI-compatible client at https://api.meta.ai/v1 and select muse-wonder-1.0. Store the key outside your source files. For an existing coding agent, follow Meta’s integration recipe so file access, shell commands and tool results are connected to the model.

Muse Code is the terminal application for working on a repository. Its pay-as-you-go mode uses API token billing; a subscription instead provides an allowance through the Muse Code credential connected during onboarding. Additional keys in the same account remain pay-as-you-go. Regional subscription benefits are displayed during signup.

For this Python example, install the openai package and set MODEL_API_KEY to your Meta key in the environment. The example asks for a repair plan; the surrounding application must execute any approved file edits or tools. Start with one reproducible issue and a known test command, then keep the response usage record with the accepted result.

The documented context window is 1,048,576 tokens. Muse Wonder accepts text, image, video, audio and PDF inputs and returns text; the model guides cover each attachment format. Put the task and evidence together, then leave room for the answer. A repository-sized window lets you supply broad context, while relevant files and concise tool results make the work easier to review.

Reasoning effort runs from minimal to max. Test a harder setting when the first attempt misses a checkable requirement. For an agent, compare total time and accepted results across the complete tool loop, including repairs. Changing Muse Wonder’s reasoning effort to max keeps the same model; Full has a separate access path.

On October 13, Meta lists the Standard-mode USD rates below for muse-wonder-1.0. There is no long-context price premium: the same rates apply as the window fills. Contributor-mode API discounts are not offered for Wonder.

For 100,000 uncached input tokens and 10,000 output tokens, the bill is $0.40 + $0.20 = $0.60. If all 100,000 input tokens hit an existing cached prefix, input costs $0.02 and the same output costs $0.20, for $0.22. Read cached_tokens in the usage response; charge uncached input at the regular rate.

Web search grounding adds $2.50 per 1,000 search queries. Four queries add $0.01 to the $0.60 token example. In an agent workflow, total the billed tokens and searches across every turn and retry. Divide that total by the tasks accepted by your review or tests.

Full is available in Muse Code on Power Usage and Maximum Usage subscriptions. Meta’s model page displays $50/month and $200/month respectively; the subscription guide says available benefits can vary by region. Full is not available through the ordinary pay-as-you-go Model API, so the $4/$20 token rates do not price Full.

To enable Full, sign in to Muse Code with your Meta Model API account and open Accounts Center’s Subscriptions section. Select Muse Code and a Power Usage or Maximum Usage plan that includes Full in your region. The subscription is attached to the Muse Code CLI credential; creating another API key does not turn it into Full access. Use /upgrade in an existing interactive Muse Code session to manage the plan.

Muse Code subscriptions default to Contributor mode, which exchanges more usage for permission to use submitted content in model training. You can turn it off in settings. Wonder and Full can use either mode in the plans that include them; Wonder’s API route uses Standard mode.

In October 12 posts reproduced by Tech Twitter and Techmeme, Theo (@theo) described using Wonder in T3 Code’s Orchestrator V2 and having Opus ask it for code, copy and architecture reviews. His follow-up clarified that the model behind that positive experience was Muse Wonder Full, and distinguished it from standard Wonder over the API.

That pairing is a workflow to try when a second review helps a repository task: keep the proposal, reviewer comments and final accepted change together. To try the same variant, use the Muse Code plan path for Full. A muse-wonder-1.0 API request selects the ordinary variant at the rates above.

Meta recommends Wonder for complex, long-running work where capability matters more than price per token, such as changes across a repository or an agent coordinating several tasks. Spark is positioned for higher-volume, latency-sensitive work and shorter tool loops. Both share the documented 1,048,576-token API window, so context capacity alone does not distinguish them.

For ordinary Standard-mode API traffic, Wonder costs $4 input and $20 output per million tokens; Spark 1.3 costs $1.25 and $4.25. The 100K-input/10K-output example is $0.60 on Wonder and $0.1675 on Spark 1.3 before cache and tools. Test Wonder when a missed requirement or repeated repair makes the cheaper attempt costly.

Spark 1.3’s documentation notes degraded audio-understanding support and directs audio work to Spark 1.2 or the separate transcription model. Wonder documents audio input alongside text, images, video and PDFs. Match the endpoint’s input support to your task instead of treating the model family as one interchangeable route.

Meta released Muse Spark Open on October 15 as an 816B-total, 42B-active open-weight model. It has its own hosted inference partners, downloadable checkpoints and Muse Model License. Wonder and Full retain their separate API and subscription access paths.

For a private deployment, Spark Open has multi-GPU recipes and an MLX path across two or four configured Mac Studios. Its Together tariff is $1.25 uncached input and $4.25 output per million tokens. Use the Spark Open guide for that model’s hardware, cache price and licensing details.

Meta’s announcement demonstrates coding and long-running agents and publishes benchmark results for Wonder and Full separately. Select the correct variant before comparing a result. For repository work, preserve the starting commit, tool permissions, task instructions and success tests so differences in the agent setup are visible.

GPT-6.1 Sol starts at $2 input and $10 output per million tokens on the standard API, with a higher tier above 272K input tokens. For the 100K-input/10K-output example, Sol costs $0.30 and Wonder $0.60 before cache and tools. Wonder keeps its ordinary input rate at longer context; compare the full request tier when the same task grows. Sonnet and Opus provide other coding-agent paths with their own cache and service rules.