▶ YouTube Summaries

← All summaries
t3dotgg#▶ Watch

Meta's Muse Wonder Full is a frontier model with a different flavor

2026-10-13 Tue ⏱ 38 min t3dotgg

Theo reviews Meta's new Muse Wonder release after several days of early API access. He argues the big version, Muse Wonder Full, is the first model from a lab other than OpenAI and Anthropic that sits in the frontier tier with Fable, Astra and Opus 5.5. It is weak at 3D and atrocious at UI, and its ceiling is a bit lower than Opus. But it rarely does dumb things, orchestrates sub-agents well, is fast and token efficient, and writes output he finds far more readable than "Claude slop". The cheaper API-only Muse Wonder is a much worse model and not worth its price. A third serious lab, he says, is good news for competition.

Sponsor: Browserbase

  • Browser infrastructure for agents: credential system integrated with 1Password so agents can sign into accounts, a search API, and a fetch API returning HTML, JSON or markdown.

The negatives first

  • Two models with confusing names: Muse Wonder (~120 TPS, API, $4 in / $20 out / $0.20 cache reads per million tokens, a bit cheaper than Opus) and Muse Wonder Full (~50 TPS, clearly bigger, only via the $50 / $200 Muse Code subscription, no published price).
  • Meta benchmarked the Full max runs (beating Fable, Astra and Opus 5.5 on many benches, state of the art on Deep Suite and SWE Marathon, strong Terminal Bench), but without a price, price-versus-performance can't be measured. Meta is reportedly compute constrained.
  • Memory was a training focus: post-training taught the model how memory works.
  • 3D is poor: Blender assets are garbage, though game movement feels decent and it beats Grok 4.7. Meta itself lists visual/spatial reasoning as a known limitation.
  • Theo's point: 3D and computer use don't emerge from raw intelligence; they come from dedicated post-training environments (likely a third-party 3D environment sold to many labs, first noticed with Kimi K3). The Opus 5.5 jump over the bigger Fable 5.1 in 3D shows this.
  • UI work is "atrocious", Sonnet 3.7-era slop. For an all-rounder go to Claude; the $200 Claude Code plan is still the best value in AI.

Skatebench oddity

  • Only Gemini models ever broke 90% on his trick-naming bench; Muse Wonder Full scores 95%. RL usually lowers this score (Opus 5 Max 84%, 5.5 Max 79%). Gemini-like skills (grammar, mental 3D transforms, vision labeling) seem baked in.

What he likes

  • A new flavor of output: concise, readable, not a Claude distillation despite Meta's large Claude usage. Default verbosity is higher; his AGENTS.md style rules (short, varied sentences, simplified technical English) make it his favorite to talk to.
  • On his "Lust" compiler project (a fast-compiling Rust subset) with 115+ agents, blind judges preferred Opus (41 vs 30 of 50 on one task). Opus does deeper, more correct work; Muse does more work faster and is a surprisingly good, fair reviewer (wrote 209 probes against Opus fixes). When the two debated plans in T3 Code, they often settled on Muse's plan with Opus feedback. Opus graded Muse B+ and itself A-.
  • Strong orchestration: well-scoped sub-agent prompts ("read actual implementation bodies for every claim you cite", "do not edit files"), good steering, handles many parallel agents. Rare outside the two big labs.
  • Low "dumbness": smart and dumb are separate axes. Astra is very smart and 10/10 dumb, Opus 10/10 smart and 3/10 dumb, Muse 8-9/10 smart and 1-2/10 dumb. It doesn't delete wrong directories or break its own state.
  • Speaker-reviewer bias caveat: as someone who tests many models, "different" is inherently appealing to him.

Real bug: remote update not updating T3 Code runtime

  • Opus found and fixed the bug but wrote too much text.
  • GPT 6.1 Soul missed that the flow was new, hit a tool error, and proposed a wrong fix (remove the npm version check).
  • Muse identified the cause (Effect HTTP client adds two trace headers to the npm check), verified it with a test fetch, and found an existing one-line PR from someone else that fixes it, in far more readable output. No benchmark measures this.

Codebase audits and PR review

  • T3 Code architecture audit: Opus 5.5 best, Sonnet 5.5 strong, then Soul, Astra, Muse Full; plain Muse Wonder about 30% below Full, close to half Opus. Many bad online demos likely use plain Wonder via API.
  • Lakebed audit: Soul best (9/10); plain Wonder 5.75. Muse Full finished in 16 minutes, about half of Soul and far less than Opus; Sonnet took 80 minutes.
  • TS Rust audit: Muse Full scored about the same as Fable 5.1, 10+ minutes faster than others.
  • Token efficiency: fewer tokens than Opus 5.5 xhigh for nearly the same score; ~54k tokens per task for both Wonder variants.
  • PR review on 8 unmerged T3 Code PRs: Opus 23 findings (21 valid, many nits, 4 novel mediums); Soul 9 findings, no nits, little found on small PRs; Muse 16 findings, most overlap with others' bugs, 5 novel, mostly low severity. Muse is a good single-reviewer middle ground.

Prioritization report

  • Asked both to deep-dive T3 Code plus issues, PRs and discussions. Opus took ~30 minutes and produced a wall of text.
  • Muse took ~17 minutes, used one critic agent to falsify claims and re-ran git counts itself. Output: clear key decisions (freeze new providers, kill legacy sidebar, treat discussions as untrusted demand), then P0 items with sizes and evidence (server stalls from a 4-worker pool, PR API quota burns of 22k GitLab calls, provider model lists, data-loss/billing repro). Much more readable, with good intuition about where to look.

Verdict

  • Not the smartest, but well above his bar and the first non-Anthropic/OpenAI model since Fable he could daily-drive; he still calls Opus for hairy work. Don't drop Claude Code for it, and pricing for Full is unknown.
  • Meta catching up makes the market more competitive, which has historically improved both intelligence and cost. Google's long-teased Gemini 4 Argan is still missing.