The Meta AI that beat Opus in the coding marathon

Create the account and key on the Meta Model API

Go to dev.meta.ai, create the developer account and generate a key. Keep it in an environment variable, never in the code.

export MODEL_API_KEY="your-key"
curl -s https://api.meta.ai/v1/models -H "Authorization: Bearer $MODEL_API_KEY"  # checks that muse-wonder-1.0 shows up

Call Wonder with the OpenAI SDK

The Chat Completions endpoint is compatible with the OpenAI format: change the base_url and the model name. For long tasks with tools, Meta recommends the Responses API (POST /v1/responses).

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.meta.ai/v1", api_key=os.environ["MODEL_API_KEY"])
r = client.chat.completions.create(
    model="muse-wonder-1.0",
    reasoning_effort="high",   # ranges from "minimal" to "max"; "max" does NOT turn it into Full
    messages=[{"role": "user", "content": "Explain what this repository does."}],
)
print(r.choices[0].message.content)

Or use it through OpenRouter

Wonder is on OpenRouter as meta/muse-wonder, at the same price (US$ 4 / US$ 20) and 1M of context. Full is not there.

client = OpenAI(base_url="https://openrouter.ai/api/v1", api_key=os.environ["OPENROUTER_API_KEY"])
client.chat.completions.create(model="meta/muse-wonder", messages=[...])

Install Muse Code (Meta's terminal agent)

It is the path to Wonder Full. Run it in the project folder, trust the folder and log in through the browser or with the key. The plans are at dev.meta.ai/docs/muse-code/subscriptions.

curl -fsSL https://dev.meta.ai/install.sh | sh        # Linux and macOS
irm https://dev.meta.ai/install.ps1 | iex              # Windows (PowerShell)
muse --version
cd my-project && muse                            # pick the Wonder model in the session
muse exec "run the tests and fix what fails"   # headless (CI), with META_API_KEY

Pick 3 to 5 real tasks, each with a test

A bug with a failing test, a medium feature, a multi-file refactor and, if you can, a long task (hours). Without a test there is no way to say who got it right.

Freeze the starting point

Same commit, a clean copy for each agent, the same request, and the verification tests out of its reach.

git worktree add ../teste-wonder <commit>
git worktree add ../teste-opus   <commit>
# same request file in both; same rules (equivalent AGENTS.md / CLAUDE.md)

Same conditions, 3 runs each

Same reasoning level, same time limit, same tools and permissions. The result varies from one run to another; three runs give an honest average.

Measure three things (and note the rest)

Wall-clock time, cost (input, cache and output tokens × price) and passing tests, including the whole suite to catch regressions. Also note how many times you had to step in.

Example row, only to show the spreadsheet format.

Review the diff blind and decide

Ask someone else (or a third model) to judge the quality without knowing who did it. Compare success rate, cost per solved task and time. Remember: Wonder is strong on long tasks and weak on pure knowledge.