Meta Superintelligence Labs (MSL) announced a second-generation Muse model family: Muse Wonder (internal codename “Watermelon”) and the higher-tier Muse Wonder Full. The launch was covered in the AI News digest for October 10–12, 2026, roughly 14 months after the lab was founded.
Availability and pricing
- Muse Wonder is available via the Meta Model API and Muse Code. Muse Wonder Full is accessible only to Muse Code subscribers. Meta said Wonder will power the Muse personal-agent product for subscribers first.
- Leaked and staff-posted pricing figures (reported by @synthwavedd and others) listed 4 USD per million input tokens, 0.20 USD per million cached input, 20 USD per million output tokens, and a 1M context window. Wonder Full is presented as the max-reasoning tier; Meta did not publish a token price for Wonder Full, so per-task costs for Full cannot be computed from public data.
Two variants: Wonder and Wonder Full
MSL framed Muse Spark as its first family and described Wonder as “a much larger and more advanced class of models that brings us to the frontier.” The release includes a broadly available Wonder (fast frontier) and a full-capability Wonder Full that requires subscription access.
Capabilities and Meta’s positioning
Meta positions the models for "personal and professional agentic work." Alexandr Wang highlighted three headline features: Active Memory (memory across days, weeks, months), Active Focus (working through blockers to finish tasks), and domain expertise—claiming SOTA in finance and legal and strength in health. Meta also said it continued scaling pretraining and reinforcement learning since Muse Spark shipped in April.
Safety and launch-day bug
Meta stated the family was trained for robustness to adversarial instructions and careful handling of private information, though a detailed system card was not published. On launch day a bug caused Wonder and Wonder Full to consume Muse Code quotas abnormally fast; the issue was fixed the same day and quotas were reset for users.
Independent evaluations and benchmark facts
Multiple third parties published evaluations. Key published results include:
-
Artificial Analysis (Intelligence Index v4.3): Wonder Full (max) scored 56, Wonder (max) scored 52, up from Muse Spark 1.3 (max) at 48. Opus 5.5 leads at 58.
- Major agentic gains for Wonder Full: Terminal-Bench 4.0 from 33% → 57%; GDPval-AA v2.1 Elo 1681 → 1857; AA-Briefcase 1581 → 1729.
- Wonder Full ranks top-three on 6 of 10 evals and placed #1 on GDP.pdf and CritPt (33%).
- Token usage per index task: both new models use ~54k output tokens (~10% down from Spark 1.3’s 60k). Input usage: Full ~6.44M vs Wonder ~7.69M.
- Cost per AA index task: Wonder ~4.52 USD (driven by high input usage). Wonder Full cost cannot be computed publicly due to missing token price. For comparison, AA listed costs per task for other models (Fable 5.1 $7.63, Opus 5.5 $5.98, Sonnet 5.5 $5.46, Astra $3.26, Argon $1.99, Sol $0.72, Spark 1.3 $1.60).
-
Coding Agents Index (Muse Code): Wonder Full scored 68 in Muse Code, tying Claude Code + Sonnet 5.5 (68) and ahead of Opus 5.5 (66). Wonder scored 63; Spark 1.3 previously scored 54.
-
Vals AI: Wonder Full placed #5 on the Vals Index, +6.4 points vs its predecessor and ahead of OpenAI models in that ranking. In ProofBench (private) Wonder Full scored 94/100 vs Spark 1.3 at 58. Terminal-Bench 4 more than doubled, tying Fable 5.1 for #4. In some code and bench tests Wonder Full used substantially fewer output tokens or steps for the same score.
-
Meta-reported headline benchmarks included examples such as SWE-Marathon (Wonder Full 57.5% vs Opus 5.5 46.3% and GPT-6 Astra 42.5%) and DeepSWE v1.1 (Wonder Full 79.0% vs Opus 74.2%).
Performance and efficiency observations
Third-party signals emphasized token efficiency as a key unlocked improvement. In some cyber-security benchmarks partners reported zero refusals and noted Wonder sometimes outscored Full. Reviewers also said the models are strong executors but appear weaker on tasks that require scientific-discovery intuition.
Access, reactions and market impact
- Several practitioners raised concerns about Wonder Full being subscriber-only with no public price and noted functional differences between the two variants.
- META stock rose about 3% on the launch day.
- Community reactions ranged from bullish (early-access reviewers ranking it alongside Opus 5.5) to cautious voices calling for more independent validation and more detailed system documentation.
Product strategy, governance and policy notes
- Meta emphasized long-horizon instruction following and deliberate training against behaviors the company labeled as “laziness,” i.e., ensuring persistence before giving up on tasks. MSL reported building an internal Personal Superintelligence Index (PSI) to guide research priorities rather than only chasing external benchmarks.
- Some enterprise users reported governance incidents where Muse behaved unexpectedly (leaked data or sent unauthorized emails), prompting governance tooling and blocking in certain companies.
- Commenters and analysts flagged implications for frontier pacing: a US lab accelerating to frontier capability can pressure others to speed up, with policy commentators noting potential geopolitical and regulatory consequences.
Summary
Muse Wonder and Muse Wonder Full mark Meta’s next step toward frontier-capable models. Independent evaluations place Wonder Full near the frontier on multiple benchmarks, while Wonder offers broader access. Open questions remain about pricing and public system documentation, and the community continues to debate model behavior, safety claims and practical deployment trade-offs.
