Independent tests put Opus 5.5 ahead of Meta's Muse Wonder
Meta's biggest model release got its first third-party scoreboard — and it's a top tier, not a top spot. Also: enterprises just funded another company to police the AI their own employees build.
Artificial Analysis has now scored Meta's Muse Wonder and Muse Wonder Full, and the independent verdict is a top tier rather than a top spot. Muse Wonder Full lands on 56 on the Intelligence Index — fourth of 689 models, tying Claude Sonnet 5.5 and clearing GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1, all at 53, but still behind Claude Opus 5.5 at 58. The API-served standard Muse Wonder scores 52, one point behind GPT-6 Astra. These are third-party evaluations rather than Meta's own charts — exactly what we said to wait for when the launch dropped, in Meta launches Muse Wonder and Muse Wonder Full models — and they mostly hold Meta's claims up while retiring the "back in the world's top three" framing. The split gets clearer in the detail: Full wins sustained agentic coding outright at 57.5 on SWE-Marathon against 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra, and 79.0 on DeepSWE 1.1, then loses badly on factual knowledge, managing about 30 on AA-Omniscience against Opus 5.5's 46 and GPT-6 Astra's 43 — a profile consistent with a smaller base model leaning on retrieval instead of knowing things. Both models take text, image, audio and video input across a million-token context window and were trained at Meta's gigawatt-scale Prometheus data center.
The catch for anyone who wants to build on it is availability and price. Muse Wonder Full has no API at all — it is locked to Muse Code subscriptions at $50 and $200 a month, with no published token pricing and no downloadable weights. Standard Muse Wonder costs $4 per million input tokens and $20 per million output, the same list rate as Opus 5.5 and roughly twice the blended cost of GPT-6 Sol or Sonnet 5.5; Artificial Analysis puts its average index task at $4.52, against $3.26 for GPT-6 Astra. Meta's model card also concedes that its comparison blends internally run evaluations, vendor-supplied scores and third-party results, and that some software engineering rows used different coding tools across vendors — the caveat VentureBeat singled out as making those rows less controlled than they appear. Meta does address the standing criticism that Muse models are tuned for benchmarks and stumble in real use, arguing Full performs better than expected on tasks outside training optimization. Read together, the independent numbers say Meta built a genuinely strong coding and agent model and a genuinely weak knowledge model, then withheld the strong one from the people most likely to test it.
Pluto Security came out of stealth with $42 million to hunt the AI agents and vibe-coded apps already inside corporate networks. The round splits into an $8.5 million seed led by TLV Partners and a $34 million Series A led by Jump Capital, with Check Point co-founder Shlomo Kramer's Skinos fund participating. Pluto's premise is a shift in what security teams watch at all: "We're moving from user behavior to agent behavior," chief executive Shahar Bahat told TechCrunch. The platform connects to the endpoint and device-management tools a company already runs — or drops in its own sensors — to inventory every agent, skill, extension and MCP server in the environment, then intervene mid-task: when an agent reaches for permissions beyond what the user intended, the system can stop the action or inject an instruction telling it to ask for confirmation first. The numbers explain the timing. Bahat says roughly 60 percent of the risks Pluto finds sit in tools built by non-developers, with finance the biggest single source ahead of marketing and legal and developers only fourth, that a newly released AI tool can appear inside a corporate environment within two hours, and that across every proof-of-concept it has run, it found something malicious — including extensions under live attack while the POC was underway.
Pluto says it started selling in April, has crossed $2 million in annual recurring revenue with more than 30 enterprise customers including Oscar Health, Workato, Sixt and AlphaSense, and employs 40 people across the United States and Israel. It also gives away Plutonium, a free checker for public AI skills, connectors and add-ons, which is where 70 percent of its pipeline now arrives from. At $42 million this is a small entry next to the $400 million rounds Island and Cyera raised on the same thesis — the bet is that enforcing behavior during a task beats reporting on it afterward.
What to watch: whether Meta opens Muse Wonder Full to the Model API, and whether other vendors start publishing their own "non-developer risk" numbers to test Pluto's claim.
Is AI governance becoming a real security category, or just this year's funding narrative? Tell us in the comments.
