3AM Marketer
AI

Meta Muse Wonder: Pricing, Specs, and Evaluation Data

Meta released Muse Wonder and Muse Wonder Full with 1M context and multimodal inputs, with API pricing reported by Latent Space at $4 per million input tokens.

$4 / 1M
5 sources, checked 14 Oct 2026.

Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, 2026, with API pricing reported by Latent Space at $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens. The release introduces two separate tiers: the standard Muse Wonder model (codename Watermelon) deployed across the Meta Model API and Muse Code, and the gated Muse Wonder Full model restricted to subscriber environments.1

What is Meta Muse Wonder?

Trained at Meta's gigawatt-scale Prometheus data center, the Muse Wonder architecture features a 1M token context window. The models process multimodal inputs across text, image, video, and audio, while restricting responses to text-only outputs. Agent execution relies on Active Memory, which carries important context across user conversations to retrieve relevant details, and Active Focus, which guides long-horizon tasks by trying alternative approaches, working through blockers, adapting to new information, requesting clarification, and checking in before high-stakes actions.12

Comparing Wonder and Wonder Full across frontier benchmarks

On the Artificial Analysis Intelligence Index v4.3, Muse Wonder Full scored 56 and Muse Wonder scored 52 at maximum reasoning effort, while Opus 5.5 reached 58. Terminal execution saw steep generational gains on Terminal-Bench 4.0: Muse Wonder Full reached 57% and standard Muse Wonder reached 52%, up from Muse Spark 1.3 at 33%. Muse Wonder Full also captured the top spot on GDP.pdf at 33% and CritPt at 33%, while taking third place on AutomationBench-AA at 71% and SciCode at 63%.1

  • Muse Spark 1.348
  • Muse Wonder52
  • Muse Wonder Full56
Artificial Analysis Intelligence Index v4.3 scores at maximum reasoning effort across frontier models in October 2026.1

In programming workflows evaluated on the Coding Agents Index inside Muse Code, Muse Wonder Full scored 68, tying Claude Code with Sonnet 5.5 at 68 and outperforming Opus 5.5 at 66, as detailed in our analysis of Claude Opus 5.5 vs Sonnet 5.5. Standard Muse Wonder scored 63 on the same index. On SWE-Marathon, Meta reports Muse Wonder Full reached 57.5% compared to Opus 5.5 at 46.3% and GPT-6 Astra at 42.5%. On DeepSWE, Meta reports Muse Wonder Full scored 79.0% over Opus at 74.2% and Astra at 74.1%, while MemoryBench showed 85.5 against Opus at 81.5, Harvey LAB showed 13.1 against Astra at 8.6, and the private ProofBench test showed 94/100 compared to 58 for Spark 1.3.1

How much does Muse Wonder cost?

At $4 per million input tokens, developers can leverage cached input discounts at $0.20 per million, similar to structures explored in Claude prompt caching. Artificial Analysis evaluation runs measured an average cost per index task of $4.52 for Muse Wonder, driven by high input token consumption. This task-level cost compares favorably to competing frontier alternatives, where Sonnet 5.5 averaged $5.46 per task and Opus 5.5 averaged $5.98 per task.1

For end users and developers using personal agent applications, the new Muse Wonder models are accessible within the Muse agent app under subscriptions of $20 or $100 per month, as reported by CNBC. By comparison, our tracking shows Claude Pro at $20 a month (or $17 billed annually) and Claude Max at $100 a month, while raw API access for OpenAI gpt-6-astra sits at $10 input and $50 output per million tokens. Muse Wonder Full has no published token price, preventing direct per-token API cost accounting.1345

Accessing the models via API and Muse Code

Muse Code installs directly with curl dev.meta.ai/install.sh | bash. While a launch-day usage-limit bug drained user quotas abnormally fast, engineering teams resolved the issue within the same day and applied a quota reset for all affected accounts.1

For teams choosing reasoning models, Muse Wonder offers a viable API target when prompt caching offsets the $4.52 task average. Teams needing top-end autonomous performance can run Muse Wonder Full inside Muse Code, but headless API pipelines require either standard Wonder or an unbundled API release for Wonder Full.1