Meta Muse Wonder: What the API Gets, and What Stays in Muse Code

1. What Meta shipped: the spec sheet

Muse Wonder is the next step after Muse Spark, the closed line Meta Superintelligence Labs started in April 2026. Meta’s launch post says both Wonder models were trained at its gigawatt-scale Prometheus data center (TechCrunch places it in central Ohio) and build on “larger, stronger base models” than Spark. Meta does not disclose parameter counts.

The two variants are split by channel, not by price tier. Muse Wonder (muse-wonder-1.0) runs on the Model API and in Muse Code. Muse Wonder Full shows up in Meta’s catalog as muse-wonder-full-1.0, but the docs say it is only available in Muse Code on the Power Usage ($50/month) and Maximum Usage ($200/month) plans, and that setting reasoning effort to “max” on the API model does not select it.

2. What improves over Muse Spark 1.3

Meta frames Wonder as the next rung on the same scaling ladder as Spark, and the API contract shows it: Wonder and Spark take the same requests on the same endpoints, so moving a workload is a model ID change. What is interesting is how far the capability moved in roughly six weeks since Spark 1.3.

On the Artificial Analysis Intelligence Index v4.3 (independent), Muse Spark 1.3 scores 48, Muse Wonder 52 and Muse Wonder Full 56. Terminal-Bench 4.0 jumps from 33% to 52% (Wonder) and 57% (Full), and Full’s hallucination rate on AA-Omniscience drops from 33% to 23%. You pay for it: list price goes from $1.25 / $4.25 to $4 / $20, and Artificial Analysis measured the cost per index task at $4.52 for Wonder against $1.60 for Spark 1.3, about 2.8x.

The three real differences

3. Benchmarks: Meta’s scorecard and the independent view

Meta’s model page compares both Wonder variants, at max reasoning effort, with Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5. The sources are mixed: some rows come from third parties (Artificial Analysis, Datacurve, Vals AI) and others from Meta’s own internal benchmarks. I kept the rows that matter most for developers and dropped Fable 5.1 for space.

The pattern is clear. Wonder leads on long-horizon coding and professional agent work, and trails Opus 5.5 on the general index, terminal tasks, from-scratch programming, health and computer use. Note how close the API model is to Full on DeepSWE (78.8 vs 79.0) and how far behind it is on the Artificial Analysis index (52 vs 56).

Artificial Analysis also scored complete coding agents. On its Coding Agents Index, Muse Code with Wonder Full gets 68, level with Claude Code on Sonnet 5.5 and above Claude Code on Opus 5.5 (66); Wonder gets 63. The weak spot is knowledge: AA-Omniscience gives Wonder 25 and Full 30, against 46 for Opus 5.5.

4. Quickstart from Meta’s docs

According to Meta’s quickstart, the Model API is drop-in compatible with the OpenAI SDK and the Anthropic SDK: base URL https://api.meta.ai/v1, a bearer key you export as MODEL_API_KEY, and three request formats (Responses, Chat Completions, Messages). Meta recommends the Responses API for agentic, multi-turn work because it can carry the model’s encrypted reasoning across turns.

Create a key in the Model API dashboard and export it. One gotcha from the docs: the OpenAI SDK does not read MODEL_API_KEY on its own, so pass it explicitly.

curl -X POST "https://api.meta.ai/v1/responses" \
  -H "Authorization: Bearer $MODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "muse-wonder-1.0",
  "reasoning": {
    "effort": "high"
  },
  "input": "Plan the upgrade of a Flask app from SQLAlchemy 1.4 to 2.0: list the breaking changes to check and the order to make them in."
}'

# pip install openai
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],  # the SDK won't read this var on its own
)

response = client.responses.create(
    model="muse-wonder-1.0",
    reasoning={"effort": "high"},  # "minimal" ... "max"; "none" returns HTTP 400
    input="Plan the upgrade of a Flask app from SQLAlchemy 1.4 to 2.0: "
          "list the breaking changes to check and the order to make them in.",
)

print(response.output_text)

# Standard mode, per 1M tokens: $4 input, $0.20 cached input, $20 output
u = response.usage
cached = u.input_tokens_details.cached_tokens   # subset of input_tokens
fresh = u.input_tokens - cached
cost = (fresh * 4.00 + cached * 0.20 + u.output_tokens * 20.00) / 1_000_000
print("reasoning tokens:", u.output_tokens_details.reasoning_tokens)
print(f"cost: USD {cost:.4f}")

export ANTHROPIC_BASE_URL="https://api.meta.ai"
export ANTHROPIC_AUTH_TOKEN="$MODEL_API_KEY"
export ANTHROPIC_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_OPUS_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_SONNET_MODEL="muse-spark-1.3"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="muse-spark-1.3"
export CLAUDE_CODE_SUBAGENT_MODEL="muse-spark-1.3"
export ENABLE_TOOL_SEARCH="true"

5. Cost math and when to pick Wonder

With reasoning models the list price is only half the story, because reasoning tokens are billed as output. Here is the arithmetic for one agent turn with 50K input tokens and 2K output tokens (reasoning included), at each vendor’s list price.

My pick, by workload

6. Honest limits

Wonder is a strong release with a few sharp edges. These are the ones I would put in front of a team before it commits.

Verdict

Muse Wonder is the first Meta model I would put on the same shortlist as Claude Opus 5.5 and GPT-6 for serious agent work. The API version is genuinely competitive on long-horizon coding and professional agent tasks, priced like Opus, faster per token, and it drops into OpenAI- or Anthropic-SDK code with a base-URL change.

It is not the overall leader: Opus 5.5 still tops the independent index and knows more, GPT-6.1 Sol hits the same index score for far less, and the variant Meta shows off most is locked inside Muse Code. My plan: route one long-horizon workload to muse-wonder-1.0 this week, log tokens and cost per task next to your current model, and keep Spark 1.3 for the cheap sub-agent tier.