Introduction
Muse Wonder 1.0 is the standard Muse Wonder model from Meta, offered by Meta as a frontier model for coding, long-running agents and computer use. This chat endpoint takes text, images, video and audio in one message list, reads up to 1,048,576 tokens of input, streams its reply and returns function calls for your application to run. Meta's larger Wonder Full variant is a separate product and is not what this endpoint serves.
| Field | Value |
|---|---|
| Input | Text |
| Output | Text |
| Input budget | 1,048,576 tokens |
| max_tokens ceiling | 1,048,576 |
| Inputs | text / images / video / audio |
Key features & highlights
Whole projects in one request
Meta positions Muse Wonder around multi-file changes in large codebases and tasks that run for hours, and the 1,048,576-token input budget is what makes that practical here. You can send the files, the failing log and the open question together instead of retrieving fragments and hoping the right ones landed. Label each file, say which version is authoritative, and state the task plainly; a large window helps with evidence, not with a vague request.
Tool calls that your code controls
The model can return several function calls in one turn, and parallel_tool_calls is on by default. Your application validates the arguments, runs each operation and replies with a tool message carrying the matching tool_call_id. That keeps the model's proposed step separate from the action you actually perform, which is the safe shape for agent loops that touch files, tickets or deployments. On this endpoint tool_choice accepts auto only; to get plain text, leave tools out.
Screens, clips and recordings as input
Messages can carry image_url, video_url, audio_url and input_audio (wav or mp3) parts next to text. Typical uses are a screenshot of a broken UI, a screen recording of a bug, or a meeting recording that needs an action list. Tell the model which attachment the question is about; unlabeled media in a long thread is the most common reason answers drift.
Dial the thinking up or down
reasoning_effort runs from minimal to max and defaults to high. Minimal or low suits classification, extraction and quick edits; xhigh and max are for hard debugging and planning. Reasoning tokens are billed as output and count toward max_tokens, so a deep setting with a small cap can spend the whole budget before the visible answer starts.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | Send Muse Wonder 1.0 an ordered conversation containing the source material for a database-library upgrade; keep tool results paired with |
max_tokens | No | For Muse Wonder 1.0, |
reasoning_effort | No | Muse Wonder 1.0 accepts minimal, low, medium, high, xhigh, max, with high as the default; compare effort settings on a database-library upgrade while monitoring output-token usage. |
stop | No | Stop Muse Wonder 1.0 with one sequence or up to four sequences; the matched marker is omitted from the final text. at least 1 characters · at least 1 items · at most 4 items |
temperature | No | Use temperature to adjust sampling in Muse Wonder 1.0; keep its default while evaluating a database-library upgrade unless you need a different answer style. |
top_p | No |
|
tools | No | Give Muse Wonder 1.0 up to 128 OpenAI-format function definitions for a database-library upgrade; your application checks arguments and performs the proposed operations. at most 128 items |
parallel_tool_calls | No | In a Muse Wonder 1.0 turn, this switch permits proposals for several independent functions; the application still controls their execution. Default: |
tool_choice | No | For Muse Wonder 1.0, |
response_format | No | Request text or structured JSON from Muse Wonder 1.0 for a database-library upgrade, then check the JSON against the fields your application expects. |
Pricing
Input $4 / Output $20 per 1 million tokens
Use cases
Multi-file refactors and reviews
Paste the relevant modules, the test output and the change request, then ask for an ordered plan with the files touched and the check that confirms each step. Because the whole context fits, the plan can reference real function names instead of guesses. Ask for a diff or a tool call and review it before anything is applied.
Long-running agent loops
Run a plan, act, observe loop where each turn adds tool results to the message list. The large window means earlier observations stay verbatim for dozens of steps. Keep a short running summary of decisions at the top so the model does not re-litigate settled choices, and stop the loop on a budget you set rather than on the model's say-so.
Bug reports from screenshots and recordings
Attach the screenshot or a short screen capture with the console log, and ask the model to state what it sees, the likely cause and the first thing to check. Separating what is visible from what is inferred makes the answer easy to verify before anyone changes code.
Document and meeting analysis
Load contracts, specs or a long call recording in a single request and ask clause-level or decision-level questions. Request quotes with locations so each claim can be traced back to the source, and treat the output as a draft for a person to confirm.
Prompt guide & pro tips
Send a chat request
POST to the chat completions endpoint with a model name and a messages array. Put instructions in a system or developer message, and use a content array when you need to mix text with image, video or audio parts. Set stream to true for long replies and read the finish reason before treating an answer as complete.
Choose effort and a safe
max_tokensStart at medium for everyday work and move up only when answers are shallow. Leave generous room in
max_tokensbecause reasoning is counted inside it; if replies cut off, raise the cap before raising the effort.Add tools and close the loop
Define tools as JSON schemas, leave
tool_choiceat auto, and when the reply containstool_callsrun each one and append a tool message with the sametool_call_id. Keep the assistant message with its tool calls in the history so the next turn has the full chain.Check cost before you scale
Billing follows measured input, cached input and output tokens, and the estimate on this page shows the expected charge for your settings. Reuse a stable prefix such as system instructions and reference files so repeated requests can benefit from cached input.
Muse Wonder 1.0 is the standard Muse Wonder model from Meta, offered by Meta as a frontier model for coding, long-running agents and computer use. This chat endpoint takes text, images, video and audio in one message list, reads up to 1,048,576 tokens of input, streams its reply and returns function calls for your application to run. Meta's larger Wonder Full variant is a separate product and is not what this endpoint serves.
Muse Wonder 1.0 can help plan a database-library upgrade: supply the current interface, the proposed version and the tests that define compatibility. Ask for a change plan before requesting a patch. This is the standard Wonder endpoint; choosing max reasoning does not switch to Wonder Full. Tool selection here is auto only, so the application must decide whether to execute each proposed operation.
Content decisions come from the selected model. Muse Wonder 1.0: NSFW behavior can vary, and no unrestricted-content guarantee is made.
API quick start
curl --fail-with-body -sN https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "meta/muse-wonder-1.0/chat",
"messages": [
{
"role": "user",
"content": "Summarise the following release notes into three bullet points."
}
],
"stream": true
}' \
| jq --unbuffered -Rrj 'select(startswith("data: ")) | .[6:] | select(. != "[DONE]") | fromjson | if .error then "\n[error] \(.error.message)\n" else (.choices[0].delta.content? // empty) end'
See the API reference for authentication, the complete schema, polling, and error handling.