Chat, coding agents & document work. Priced per 1M tokens — input, cached input and output are billed separately. GPTProto is 20% below official rates.
| Scenario | Meta list | OpenRouter | GPTProto | You save / mo |
|---|---|---|---|---|
| Personal10M tokens / mo (4.8M cached) | $53.76 | $56.72 | $43.01 | −$10.75≈ $129.02 / yr |
| Team100M tokens / mo (48M cached) | $537.60 | $567.17 | $430.08 | −$107.52≈ $1290.24 / yr |
| Business500M tokens / mo (240M cached) | $2688.00 | $2835.84 | $2150.40 | −$537.60≈ $6451.20 / yr |
Meta · $3.2/M input · $16/M output
Bytedance · 262K context · $0.0971/M input · $0.9714/M output
Bytedance · 262K context · $0.0971/M input · $0.9714/M output
Build with Meta’s reasoning model for demanding coding and agent tasks. Muse Wonder accepts text prompts and returns text, with a native 1M-token context window, adjustable reasoning effort, and tool calling. Use it to turn detailed requirements into code, explanations, or implementation plans. Meta
Coding Across Multiple Files
Work through changes that span application logic, tests, and configuration. Muse Wonder is designed for repository-level engineering, including planning edits and revising an approach after test feedback. research.meta.ai
1M-Token Context Window
Work with extensive source code, specifications, conversation history, and retrieved text in a 1,048,576-token context window. Keep related material together when a task depends on details spread across many inputs. Meta Model API
Tool Calling for Agent Workflows
Connect reasoning to your application’s functions. Muse Wonder supports tool calling and parallel calls; your agent runs the tools and returns results so the model can plan the next action. Meta Model API
Adjustable Reasoning Effort
Tune reasoning depth to the task. Meta supports effort levels from minimal to max; higher effort can improve difficult problem solving while increasing reasoning tokens, response time, and cost.
Muse Wonder is Meta’s model for complex software engineering and sustained agent work, introduced on October 12, 2026. It is available through Meta Model API as muse-wonder-1.0 and through Muse Code. Meta currently lists the API model in preview.
For text-to-text tasks, it can work from instructions, source code, error messages, and retrieved passages to produce answers, code, and task plans. It is a candidate for work that requires tracking requirements and revising an approach across several steps.
Your application supplies the conversation history, tools, and execution environment. Calling the model alone does not create a coding workspace, browser controller, or persistent memory service.
Generate implementations from specifications, investigate stack traces, review changes, and propose refactors. Give the model relevant code and test feedback so later responses can address what actually failed.
Compare architecture options, work through constraints, and explain a proposed decision. Include the assumptions and acceptance criteria that the answer needs to satisfy. Increase reasoning effort when a task requires deeper analysis. Meta Model API
Use tool calling to retrieve records, query services, or trigger defined actions. The model selects a function and arguments; your application executes the function and returns its result. Meta Model API
Where schema-based output is supported, turn supplied text into consistent JSON for ticket fields, categories, or configurations. Validate factual accuracy as well as the response format before passing results downstream.
Muse Wonder is the API-accessible model represented by this page. Wonder Full is a separate variant; setting Wonder’s reasoning effort to max does not switch to Full.
| Decision factor | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Official API access | Available through Meta Model API | Currently unavailable through Meta Model API |
| Muse Code access | Pay-as-you-go and subscription plans | Power Usage and Maximum Usage subscriptions |
| Model selection | muse-wonder-1.0 |
Selected through eligible Muse Code access |
| Choose it when | Building your own coding assistant or agent | Using Meta’s higher-capability variant inside Muse Code |
Muse Wonder suits difficult tasks that need sustained reasoning and repeated feedback. For a workflow with several stages, evaluate it on planning, implementation, or the hardest reasoning step, then compare cheaper models for routine classification and short summaries. This allocation is a practical starting point based on Meta’s workload guidance.
Measure the complete task: accepted output, total tokens, time taken, and retries. Higher reasoning effort consumes more billed output tokens, including internal reasoning, so a short visible answer can still have a substantial token cost.
GPTProto lets you evaluate supported models with one key and a shared balance. Explore LLM API for AI Agents for workflows that assign different models to different stages.
Guides, comparisons, and updates related to this model.
All Articles
Compare 11 Together AI alternatives for hosted inference, GPU rental, and fine-tuning. Review pricing models, catalog coverage, and migration effort.
Compare 10 CometAPI alternatives for unified access to GPT, Claude, Gemini, and media models. Explore pricing, model coverage, free tiers, and migration.
Claude Haiku 5.5 cuts input costs to $0.10 per million tokens — until your prompt passes 100K and the whole request bills at 5x. Pricing, benchmarks, API changes, and how it compares to Sonnet 5.5 and GPT-6 Luna.
