Muse Wonder

Original Article ↗ Hacker News Discussion ↗

Meta’s new Muse Wonder large language model is being received as a near–frontier system, with benchmark scores reportedly just behind Anthropic’s Opus 5.5 and ahead of several other top-tier models. Commenters debate whether this signals that AI capabilities are rapidly commoditizing—favoring deep-pocketed tech giants with massive compute and data—while also questioning benchmark overfitting, real-world agent reliability, and Meta’s weak privacy track record. Many welcome stronger competition to OpenAI and Anthropic, but reservations remain over Meta’s pricing, opaque access to the highest-end “Wonder Full” variant, and the broader concentration of power among a handful of AI labs.

Model performance & frontier status

  • Muse Wonder / Wonder Full are presented as near-frontier; Meta claims Wonder Full is just behind Opus 5.5 on Arena-style benchmarks.
  • Commenters report mixed real-world performance: some see Wonder roughly at Sol 6.1 level, others say it feels closer to Luna / Haiku / Sonnet depending on task.
  • Vision appears weaker: HTML-from-image tests place Wonder between Haiku and Sonnet, well behind Opus 5.5.
  • Coding reviews are split: some find Muse Spark already a strong daily driver; others say Wonder still trails Astra / Opus 5.5 / Fable on complex dev workflows and can get stuck in loops.

Benchmarks vs reality

  • Many assume all major labs “benchmaxx” by training on benchmark data; some note Meta explicitly discusses this in its post.
  • Arena rankings for Wonder fluctuated (briefly #2, then lower), reinforcing skepticism about headline numbers.
  • Several users stress that benchmarks are diverging from day-to-day usefulness (planning trips, agents, coding harness reliability).

Pricing, subsidies, and access

  • Wonder API pricing ($4 in / $20 out per million tokens) is comparable to top-tier rivals and above some competitors.
  • Muse Code subscriptions: Spark 1.3 tiers are seen as extremely subsidized (orders of magnitude of value), but Wonder initially burned quotas very quickly; later described as a bug.
  • Some complain basic/entry plans are too limited; others praise generous third‑party bundles (e.g., OpenCode Go) using cheaper models.
  • Wonder Full is restricted to Meta’s own higher-tier subscriptions/harness; no general API price given.

Open weights, data, and commoditization

  • Commenters doubt Wonder weights will be released; frustration that earlier open-weights promises (for Spark) are still unfulfilled.
  • Debate on whether LLMs are commoditizing:
    • One side: if enough money + compute can buy frontier capabilities, models become commodities and only a few nations/giants will own them.
    • Other side: even if models commoditize, moats may come from data, infrastructure, distribution, or ecosystems.

Meta’s strategy and competition

  • Some see Meta’s huge AI CapEx, acquihires, and use of contributor traffic (tens of trillions of tokens) as finally paying off, placing it alongside OpenAI and Anthropic at the frontier.
  • Others argue it’s easier to follow leaders than to lead; frontier leadership is rotating rapidly (OpenAI, Anthropic, Google, xAI, DeepSeek, now Meta).
  • There is debate whether this shows “great execution” or simply brute-force compute plus platform leverage.

Product experience & agents

  • Muse (consumer agent) is praised as a major launch and easier on‑ramp for “normies” via Facebook/Instagram/WhatsApp, but:
    • Some report poor reliability on real-world tasks (travel/restaurant planning, up-to-date local info).
    • Muse Code’s sandbox and quotas are criticized (bugs on some OSes, high token burn, confusing usage limits).
  • Confusion over model tier naming (Spark, Wonder, Wonder Full) and which versions power the assistant vs code product vs API.

Trust, privacy, and brand

  • Strong distrust of Meta’s privacy record; some refuse to use Muse “as much as” they’d trust it with intimate data (i.e., not at all).
  • One user initially reported seeing other people’s accounts in the login UI; later traced to an interaction with a content blocker, but the episode reinforced concerns about tracking and fingerprinting.
  • Broken unsubscribe flows and cross‑product data sharing (e.g., showing friends as “already on Muse”) are cited as signals that user privacy is a low internal priority.

Safety, governance, and “superintelligence”

  • Heated discussion about the term “superintelligence”:
    • Some object to its political rebranding; others note Meta used it earlier as lab branding.
  • Broader debate on recursive self‑improvement:
    • Pro‑RSI: rapid capability scaling so far, heavy capital chasing AGI, many measurable evals.
    • Skeptical: verification & alignment are hard, benchmarks are gameable, intelligence gains may slow, and RSI may optimize for measurable skills while neglecting safety.
  • Opinions differ whether multiple competing frontier labs are safer (mutual checks) or more dangerous (race dynamics, cartel risks).