Muse Wonder and Wonder Full: Pricing, Benchmarks, API Access and Data Modes

Meta now has a frontier model it does not give away. On October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, which Meta calls "our most capable models to date." Unlike Muse Glimmer and Muse Spark Open, neither model has downloadable weights. Meta says they build "on the groundwork we laid with Muse Spark," and VentureBeat reports both were developed at Meta's Prometheus AI data center.

For engineering teams, the main detail is which model you can actually reach. Muse Wonder runs on the pay-as-you-go Meta Model API, priced exactly like Claude Opus 5.5. Muse Wonder Full, the one at the top of Meta's own charts, is only available inside a Muse Code subscription. If you are building a product on an API, you get Wonder, not Wonder Full.

This guide covers pricing and cost math, the API and data-mode rules (Standard, Contributor and ZDR), rate limits, the documented setup for running Claude Code and Codex on Muse Wonder, and how the benchmarks compare once you separate Meta's table from independent data. All figures are as of October 2026.

Muse Wonder vs Muse Wonder Full at a glance

Meta's models page says plainly that setting reasoning effort to "max" on muse-wonder-1.0 does not give you Wonder Full. They are separate models, not one model at two effort settings.

What Muse Wonder is built to do

Meta pitches Wonder at "personal and professional agentic tasks," naming law, finance, health and frontier coding as target domains. The launch post highlights two behaviors:

• Active Memory. Meta describes this as "the ability to carry what matters to you across conversations." The model brings back earlier details when they become relevant to the current task. Meta measures it with MemoryBench, an internal benchmark of 64K to 100K-token multi-turn conversations with distractor and conflicting turns.
• Active Focus. The model keeps working toward a goal when a task gets hard. It tries other approaches, recovers from errors, asks for clarification and checks in before important actions. Meta evaluates it with FocusBench (internal) and SWE-Marathon, a long-running software-project benchmark.

Meta's models page lists these API features for muse-wonder-1.0: chat completion, image understanding, video and audio understanding, tool calling, structured output, search grounding and computer use.

Muse Wonder pricing (as of October 2026)

Some details from Meta's pricing page:

• No long-context surcharge. You pay the same per-token rate whether the window is nearly empty or almost full.
• Web search costs $2.50 per 1,000 queries, on top of token costs.
• No Contributor discount. "Contributor mode isn't available for Muse Wonder on Model API." Muse Spark offers Contributor pricing as low as $0.10 input and $0.20 output, but Wonder does not.

Worked example: an agent workload

Take a coding agent that sends 1,000 requests a day. Each request carries 50,000 input tokens, of which 40,000 are a cached prefix (system prompt, repo map, tool definitions), and returns 2,000 output tokens.

• Uncached input: 10,000 × $4.00 / 1,000,000 = $0.040
• Cached input: 40,000 × $0.20 / 1,000,000 = $0.008
• Output: 2,000 × $20.00 / 1,000,000 = $0.040
• Per request: $0.088. Per day: $88. Per 30-day month: $2,640.

This example assumes a warm cache and leaves out cache-write charges. Meta's pricing page lists only input, cached input and output rates, with no separate cache-write line, but AA's run (below) shows cache writes can be the largest single slice of Wonder's input spend, so measure your real bill during a pilot.

If the same agent also runs 200 web searches a day, that adds 200 × $2.50 / 1,000 = $0.50 per day, or $15 a month. Without caching, the 40,000-token prefix would cost $0.16 per request instead of $0.008. That raises the per-request total to $0.24 and the monthly bill to $7,200. On this model, prompt caching is not optional.

What a real task costs: Artificial Analysis data

Artificial Analysis (AA) measured the average cost of one Intelligence Index task in its launch analysis:

AA reports Wonder used 7.69M input tokens and about 54K output tokens per task. Output cost $1.08 (54,000 × $20 / 1M = $1.08) and input $3.43 (76% of the total), split into $1.92 for cache writes, $1.44 for cache hits and only $0.07 for uncached input. For comparison, if all 7.69M input tokens had been billed at the $4.00 uncached rate, input alone would cost about $30.76 (7.69 × $4.00). Agent loops re-send a lot of context, and caching is what keeps Wonder cheaper per task than Opus 5.5 in AA's launch figures despite the identical list price.

These per-task figures move. When we checked AA's live Muse Wonder model page on October 26, 2026, it showed $8.63 per Intelligence Index task, well above the launch analysis. Use the launch table for like-for-like comparisons between models, and re-check AA before you budget on it.

On speed, AA's model page showed Wonder (max) at about 197 output tokens per second on Meta's API, with a time to first token of about 36.5 seconds, when we checked on October 26, 2026. Plan for latency on interactive surfaces.

API access, data modes and ZDR

The Meta Model API accepts OpenAI SDK code against https://api.meta.ai/v1 (Responses API and Chat Completions). Anthropic-format clients use the Messages API at https://api.meta.ai. Meta's quickstart notes that the OpenAI SDK does not read MODEL_API_KEY automatically, so pass it explicitly:

The sample below is adapted from Meta's quickstart, which uses muse-spark-1.3 and no effort setting; we swapped in Wonder and added a reasoning effort.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)

response = client.responses.create(
    model="muse-wonder-1.0",
    input="Review this migration plan and list the riskiest steps.",
    reasoning={"effort": "high"},  # "minimal" to "max" per Meta's models docs
)

print(response.output_text)

Standard, Contributor and ZDR

The ZDR limits affect architecture. If you need ZDR for compliance, you cannot use server-side conversation state, background responses, file uploads or the built-in web_search tool. You manage conversation state on the client, upload nothing, and bring your own search tool. Meta's terms also allow retention for legal compliance and misuse investigations, according to VentureBeat's reading of them.

Contributor mode is a Muse Code issue, not an API issue. Muse Code changelog 1.4.6 says "Contributor mode is now on by default for subscribers." The changelog does not name Wonder Full, but since Wonder Full is only available through those subscriptions (and Meta's models page lists both Wonder variants in Standard and Contributor mode in Muse Code), assume Contributor mode applies unless you turn it off in /settings. Meta's terms prohibit submitting confidential, sensitive or personal information under Contributor services. We cover the tradeoff, and how to switch it off, in our Muse Code Contributor-tier privacy guide.

Rate limits

Meta's pricing page lists one set of limits for Muse Wonder, applied per team rather than per API key:

• 3,000 requests per minute
• 4,000,000 tokens per minute
• 600 background response submissions per minute per team

The token cap binds well before the request cap on long-context work. At 4M TPM, you can send about four near-full 1M-token requests per minute across your entire organization. For the 50K-token agent requests in the example above, that works out to 80 requests per minute (4,000,000 / 50,000). Batch-heavy pipelines should shard across teams or keep prompts lean.

Wonder Full and Muse Code plans

Muse Wonder Full is only available through Muse Code, Meta's coding agent. Meta's Muse Code product page lists four plans:

All plans have usage limits, and Meta has not announced a per-token price for Wonder Full. In practice, Wonder Full is an interactive tool for individual engineers, not something you can build a product on. If your evaluation shows you need Wonder Full quality inside your own service, today you cannot get it. Decide whether Wonder, Opus 5.5 or GPT-6 Astra is good enough for your workload. For a closer look at the agent itself, see Muse Code vs Claude Code.

Running Claude Code and Codex on Muse Wonder

Meta documents official setups for running Claude Code and OpenAI's Codex CLI against muse-wonder-1.0. Both use a pay-as-you-go Model API key, so you are billed at the per-token rates above, in Standard mode.

Claude Code

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"
claude

Meta's caveats:

• The base URL has no /v1, because Claude Code appends /v1/messages.
• Use ANTHROPIC_AUTH_TOKEN (bearer token), not ANTHROPIC_API_KEY (which sends an x-api-key header).
• Point every model alias at a Model API model. Otherwise Plan Mode, subagents or background tasks may try to call a Claude model that Meta doesn't serve. To run everything on Wonder, set all five model variables to muse-wonder-1.0.
• ENABLE_TOOL_SEARCH keeps MCP tool search on, because Claude Code disables it for non-first-party hosts by default.

The default split above sends subagents and background work to Muse Spark 1.3 at $1.25/$4.25, which keeps costs down. Only the main loop runs on Wonder.

Codex

# ~/.codex/config.toml
model = "muse-wonder-1.0"
model_provider = "meta"
model_reasoning_effort = "high"           # none | minimal | low | medium | high | xhigh | max
model_reasoning_summary = "auto"
model_context_window = 1048576            # 1M-token context
model_supports_reasoning_summaries = true
model_auto_compact_token_limit = 900000

[model_providers.meta]
name = "Meta Model API"
base_url = "https://api.meta.ai/v1"
env_key = "MODEL_API_KEY"
wire_api = "responses"

Meta's models page lists effort from "minimal" to "max"; the Codex config comment above also lists "none", so test that value before relying on it.

Export MODEL_API_KEY and run codex (or codex exec "..." for non-interactive runs). Meta notes that wire_api = "responses" is required so reasoning carries across turns. If you're signed into Codex with an OpenAI account, run /logout first. To keep this setup separate from an existing install, set CODEX_HOME to a separate directory.

Benchmarks: Meta's table vs independent data

Meta's model card compares both Wonder variants against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5, all at max reasoning effort. Here are the coding rows plus a few agent rows (Meta's table, from the Muse Wonder model card):

A few things to note when reading this table:

• Terminal-Bench numbers differ across our posts. Opus 5.5 shows 59.6 here but 66.4 in our Gemini 4 Argon and Kimi K3.1 Pro guides. Meta's row comes from Artificial Analysis's independent run on AA's own harness. Those other posts use vendor-reported Terminal-Bench 4.0 tables, where each lab runs its own agent setup. Different harnesses give different scores for the same model, so don't compare the two figures directly.
• Wonder vs Wonder Full on DeepSWE is close. 78.8 vs 79.0 means the API model gives up almost nothing on long-horizon repo work by that measure. The gap is wider on Terminal-Bench 4.0-AA (52.0 vs 57.1) and ProgramBench (50.0 vs 59.0).
• Meta notes its own caveats. The SWE-Marathon row mixes harnesses (Codex for Astra, Claude Code for Fable), and Wonder Full ran at xhigh there. Competitor runs Meta generated internally are "best-effort" and "may not be optimally tuned for proprietary models."

Independent: Artificial Analysis

AA's launch analysis puts Wonder Full at 56 and Wonder at 52 on Intelligence Index v4.3, up from 48 for Muse Spark 1.3. Claude Opus 5.5 (max with fallback) still leads AA's leaderboard at 58. Other AA results for the Muse models:

• Coding Agents Index (scores agent and model pairs): Muse Code on Wonder Full 68, tied with Claude Code on Sonnet 5.5 (68) and ahead of Claude Code on Opus 5.5 (66); Muse Wonder 63 (AA does not name the harness for this entry); Muse Code on Muse Spark 1.3 54.
• GDPval-AA v2.1 (Elo): Wonder Full 1857 vs Opus 5.5 1866 and Sonnet 5.5 1838.
• AA-Omniscience: Wonder Full 30, Wonder 25, well behind Opus 5.5 (46), GPT-6 Astra (43) and Gemini 4 Argon (42).
• Hallucination rate: Wonder Full 23%, Wonder 27%, vs Gemini 4 Argon (high) at 15%. AA credits Wonder Full's 5-point AA-Omniscience edge over Muse Spark 1.3 (30 vs 25) to a lower hallucination rate, 33% down to 23%.

Overall, Wonder is a credible frontier coding model that matches Opus 5.5's list price and cost less per task in AA's launch analysis. It trails Opus 5.5 and Astra on factual recall and hallucination. If your workload depends on knowledge rather than agentic execution, test that gap before you switch. For more on the Opus side, see our Claude Opus 5.5 guide.

Safety and limitations

Meta's model card reports a Meta Superintelligence Scaling Framework assessment run on Wonder Full. Without mitigations, Meta does not rule out the high-risk threshold in the chemical and biological domain, so it applied safety post-training and system-level classifiers. With those in place, Meta rates the model "moderate or lower" across all domains. On October 23, Meta corrected an ExploitGym cyber result after fixing its harness: Wonder Full captured 20 intended flags, not the 21 originally reported. The launch post still says "a full report with additional safety evaluations will be available soon."

Meta's listed limitations include errors in multi-step reasoning on novel tasks, weak visual and spatial (especially 3D) reasoning, and performance that varies with "prompts, tools, and the agent harness." Meta recommends human-in-the-loop confirmation for irreversible actions in agentic deployments. Build that in for any agent with write access to production systems.

Should your team use Muse Wonder?

• Yes, worth evaluating if you run long-horizon coding agents, legal or finance document workflows, or multimodal agents and want a second frontier vendor at Opus-level list pricing. The documented Claude Code and Codex setups make it easy to A/B test inside tools your team already uses.
• Wait if you need Wonder Full quality through an API, need Contributor-mode pricing on the API, or operate in South Korea.
• Check carefully if you need ZDR, because it removes server-side state, background requests, file uploads and built-in web search. Also check if your workload depends on factual recall, where AA scores Wonder well below Opus 5.5.

If you are coming from the Spark family, our Muse Spark 1.3 guide covers the cheaper model that Wonder's default Claude Code setup uses for subagents.

FAQ

How much does Muse Wonder cost?

As of October 2026, muse-wonder-1.0 costs $4.00 per 1M input tokens, $0.20 per 1M cached input tokens and $20.00 per 1M output tokens on the Meta Model API, with no long-context surcharge. Web search adds $2.50 per 1,000 queries.

Can I use Muse Wonder Full through the API?

No. Muse Wonder Full is only available in Muse Code on the Power Usage ($50/month) and Maximum Usage ($200/month) plans. Setting reasoning effort to "max" on Muse Wonder does not route to Wonder Full.

Does Meta train on my Muse Wonder API data?

Not on the pay-as-you-go API, which runs only in Standard mode, where content is not used for training. Muse Code subscriptions default to Contributor mode, which does allow training unless you turn it off.

What is the context window and rate limit?

1,048,576 tokens. Rate limits are 3,000 requests and 4,000,000 tokens per minute per team, with background submissions capped at 600 per minute.

Can I run Claude Code on Muse Wonder?

Yes. Meta documents it: set ANTHROPIC_BASE_URL=https://api.meta.ai, pass your key as ANTHROPIC_AUTH_TOKEN, and point ANTHROPIC_MODEL and the other model aliases at Model API models such as muse-wonder-1.0.

Is Muse Wonder better than Claude Opus 5.5?

It depends on the benchmark. On Meta's table, Wonder leads Opus 5.5 on DeepSWE 1.1 (78.8 vs 74.2) but trails on Terminal-Bench 4.0-AA (52.0 vs 59.6). On AA's Intelligence Index v4.3, Opus 5.5 scores 58 to Wonder's 52, but Wonder cost less per task in AA's launch analysis ($4.52 vs $5.98).

Sources

• Meta Newsroom: Introducing Muse Wonder models
• Meta AI Research: Muse Wonder
• Meta: Muse Wonder Model Card (benchmark table, methodology, risk assessment)
• Meta Model API docs: Models
• Meta Model API docs: Pricing and rate limits
• Meta Model API product page
• Muse Code product page (plans)
• Meta Model API docs: Coding agents (Claude Code, Codex)
• Meta Model API docs: Quickstart
• Muse Code changelog
• Artificial Analysis: Muse Wonder analysis
• Artificial Analysis: Muse Wonder model page
• Artificial Analysis: Model leaderboard
• VentureBeat: Meta's new Muse Wonder and Muse Wonder Full models

Pointing Claude Code or Codex at a new endpoint is the easy part. Building agentic integrations that handle ZDR constraints, caching and human checkpoints takes engineers who have done it before, and Codersera can help you hire vetted remote AI developers for that work.