Meta's Muse Wonder Launch Creates Buying Decision Dilemma for Developers

Meta has introduced its new reasoning model, Muse Wonder, which offers developers a serious alternative to existing agents that can work across code, documents, and applications. However, this launch also creates a buying decision dilemma: whether to use the US$4-input/US$20-output API model or pay for access to the stronger Muse Wonder Full inside Muse Code.

The evidence suggests that Wonder Full is strongest on several measures of sustained agent work and long-running software engineering. Claude Opus 5.5 still leads other coding and general-intelligence comparisons, but regular Wonder gives developers a public API and a million-token context window at a lower price point than Opus 5.5.

However, the token prices for regular Wonder match those of Opus 5.5 rather than undercutting them. The right choice depends on the task requirements, access needs, and total bill. Developers must weigh these factors to make an informed decision about which model to use.

Muse Wonder and Muse Wonder Full are reasoning models aimed at personal and professional agents: software changes, research, document work, and tasks that require repeated interactions with tools. Wonder Full is a separate, higher-capability variant of the regular Wonder API, but selecting maximum reasoning on the regular Wonder API does not turn it into Full.

The pricing for Muse Wonder is based on direct API charges, which include US$4 per million uncached input tokens, US$0.20 per million cached input tokens, and US$20 per million output tokens. There are no long-context premium rates listed by Meta, but built-in web search adds an additional charge of US$2.50 per 1,000 queries plus the model’s token charges.

Wonder is also listed in OpenRouter’s catalog as meta/muse-wonder with matching API rates to those on Meta’s website. However, provider routing, platform fees, and account access should be verified at checkout. The usage descriptions for Wonder Full do not promise a fixed number of maximum-effort tasks, and subscription credentials work through Muse Code.

A separate rollout is underway in the Muse personal-agent app, where Wonder will power experiences soon with subscribers first. However, this statement does not confirm Wonder access for every current app user, as model availability and subscription benefits can vary by region. Meta’s geographic policy does not guarantee worldwide access; check the onboarding screen for your account and country.

The organizational context of Muse Wonder is covered in a separate article from September 28, which explains the Meta Enterprise Platform offering, including the Muse agent, Business Agent, model API, and Muse Code. This broader enterprise offering is separate from the Wonder-specific model access and pricing above.

An agent’s bill also depends on how often it rereads context, how long it reasons, which tools it calls, and how many attempts it needs. Meta’s reasoning documentation confirms that internal reasoning tokens count as billed output, even if a short visible answer follows a larger quantity of paid reasoning.

The following example holds token usage constant: 100,000 uncached input tokens and 20,000 billable output tokens, including reasoning. These are calculated prices, not measured completion costs, using standard direct-API rates. Search, tax, retries, cache writes, storage, and premium service tiers are excluded from these calculations.

Rate sources include Meta, Anthropic, OpenAI, Google, and DeepSeek. However, local clock time alone is insufficient to identify the applicable rate; weekday peak windows of 01:00–04:00 and 06:00–10:00 UTC should be considered for some providers. Large context still has a cost even without a premium rate.

One million uncached input tokens plus 20,000 output tokens would cost US$4.40 on Wonder. Multiply context processing across repeated turns or subagents, and the bill can grow quickly. The practical metric is the cost of a correctly completed job, including failed attempts and human repair.

Independent benchmark accounting reinforces this point. Artificial Analysis reports an estimated weighted cost per Intelligence Index task of about US$4.52 for Wonder at max effort. Its contemporaneous data gives approximately US$3.26 for GPT-6 Astra and US$5.98 for Opus 5.5 at their listed settings.

We followed the small coding pilot with a longer requirements-memory test on October 12, 2026. The test used two variants of one fictional workflow: full-history answers reproduced 240/240 final configuration facts; handoff answers reproduced 288/288 facts after a fresh conversation received only a model-written note.

The follow-up tested remembering the rules and applying them are separate tests. Each scenario began with five campaign-routing profiles and 48 checked configuration facts, including partial reversions, profile rename, retirement of old alias, pause, fallback rules, replacement exclusions, and correction.

Before the final answer was generated, we saved the source conversation. One branch reconstructed the configuration from that full history; a separate branch asked the same model to write a memory note, then gave only that note and original output/decision contract to a fresh API conversation.

The correct configuration does not guarantee a usable answer. Among scorable answers, the final full-history branches answered 49/50 proposal decisions correctly; handoff branches answered 59/60. Across all 24 audits, 19 were valid strict JSON responses.

A scoring amendment was made after observing malformed overall JSON and trailing prose in Claude baseline answers. The original strict grades remain in evidence. A post-observation change separated response-format compliance from recoverable requirement content.

The follow-up used an estimated US$1.1680 in tokens. Including the first pilot, the total was US$2.0762. These are estimates from returned usage, not attributable invoices. No additional subscription, paid judge, or generation retry was used.

We used the same three native model IDs at medium effort: muse-wonder-1.0, gpt-6.1-sol, and claude-sonnet-5-5. Each had a 4,096-output-token request ceiling and sampling settings left unset. Tools were disabled; one standard repair was available per failing checkpoint.

The original 40,000 cumulative input-token allowance stopped Claude’s first run before the unsent recall question, after all its coding checkpoints had passed. We raised that allowance equally to 60,000 for every model and resumed only that recall.

Every model received the complete visible conversation; native tokenizers, reasoning budgets, and caching behavior differ among models. Medium does not mean equal compute; Claude requests originated from a different host to keep its credential there.

Wonder retained the tested requirements in this short task. GPT did too at a lower estimated token cost in this setup. Claude also passed throughout. However, we cannot name a coding-quality winner from six perfect runs on one small project.

This test did not exercise native memory or compaction features; it only checked recall and application of rules. Repeated edge-case checks are useful but do not become independent projects just because there are many of them.