Meta Muse Wonder Guide

Muse Wonder is Meta’s newest family of AI models, launched on October 12, 2026. It is built for agent work: changing software, research, working with documents, and tasks that need many steps with tools.

It is the successor to Muse Spark, the model behind the Muse personal agent. Like Muse Spark, it is closed-weight, unlike Meta’s older Llama models.

Muse is the agent. Muse Wonder is the brain Meta is putting behind it.

Muse Wonder

The standard model, available through the Meta Model API and Muse Code. Its model ID is muse-wonder-1.0.

Muse Wonder Full

Meta’s most capable option. At launch it is only in Muse Code, on eligible Power and Maximum plans, with no public API price.

Muse Code

Meta’s coding agent app, out of beta since August 31, 2026

Meta Model API

For developers building their own apps and agents

Muse app

Meta says Muse Wonder will power the personal Muse agent, rolling out first to subscribers

Rollouts can differ by account and country. Check what your app or account shows.

Muse Wonder

$4 per million input tokens, $0.20 cached input, $20 output

Muse Spark 1.3 Standard

$1.25 input, $0.15 cached input, $4.25 output

Muse Wonder costs about three to five times more than Spark 1.3. Use Spark for simple, high-volume work and Wonder where quality matters. Prices are as of launch, so check Meta’s pricing page.

Active Memory

Pulls relevant context from earlier conversations

Active Focus

Helps the model keep going on hard tasks, adapt to changes, and ask you for input

Agentic reasoning

Plans and uses tools across long, multi-step tasks

On Meta’s own benchmark chart, Muse Wonder Full leads on some long coding benchmarks, such as DeepSWE and SWE-Marathon, while Claude Opus 5.5 leads on others, such as ProgramBench. Some comparisons used different agent setups.

No benchmark knows your work. Test on your own tasks before switching.

• Download Muse Code only from Meta’s official site
• Start in an empty practice folder
• Choose Muse Wonder as the model and set the reasoning effort
• Start with read-only permissions
• Check your settings, model, and usage allowance in the app
• Give it a small task with clear rules

• Pick three real tasks you already know the answers to
• Run them on Muse Wonder and on the model you use now
• Check facts, maths, and code yourself
• Change a rule mid-task and see if it adapts
• Give it an impossible request and see if it says so
• Compare quality, speed, and cost

• Muse Code subscriptions default to Contributor mode, which lets Meta train on your prompts and responses. Turn it off in settings if you prefer
• Muse Code runs on your computer but sends work to Meta’s servers
• Do not add passwords, keys, or private client data to prompts
• Test with non-sensitive material first
• Review exported transcripts before sharing them

• Keep approvals on for file changes and commands
• Limit network access for agent tasks
• Pressing Esc stops the current turn, but check for background tasks and subagents too
• Confirm whether you are billed through a subscription or an API key before long runs

• Fixing bugs and building features in a codebase
• Research with sources
• Turning long documents into summaries and plans
• Multi-step agent workflows with tools
• Planning tasks with clear rules and budgets

You are helping me with [task]. Rules: [fixed rules, limits, budget, deadlines]. Use only these files or sources: [list]. Work step by step, show your plan before acting, and cite sources for every fact. If any rule cannot be met, stop and tell me instead of working around it. Finish with: what you did, what you are unsure about, and what needs my decision.