Published 14 October 2026 · Updated 19 October 2026
Meta has made the jump most people assumed was still years away. Meta Superintelligence Labs released its long-anticipated Muse Wonder models this week, and the flagship Wonder Full has landed joint second in Artificial Analysis’ intelligence rankings, level with Anthropic’s Claude Sonnet 5.5. Only Anthropic’s Opus 5.5 sits above it.
For most of this cycle, the frontier has been a two-horse race between OpenAI and Anthropic. Meta’s arrival changes the shape of that contest, and not just on benchmarks. The models are strong, but the infrastructure wrapped around them may be the more worrying part for rival labs.
The numbers put Wonder Full in the top tier
Artificial Analysis’ Intelligence Index gives Wonder Full a score of 56. Google’s Gemini 4 Argon, OpenAI’s GPT-6 Astra, and Anthropic’s Claude Fable 5.1 each score 53, so Wonder Full is three points clear of that group, level with Claude Sonnet 5.5, and two points behind Opus 5.5 at 58 (VentureBeat).
A three point lead over three of the most capable models in existence is not a rounding error. It places Meta’s model in genuine frontier territory, a statement the company could not make with confidence even six months ago.
Where it wins, and where it loses
The published comparisons show a familiar pattern of specialisation. Wonder Full leads Opus 5.5 and Astra on legal work (13.1 on Harvey’s LAB-AA legal benchmark against 4.2 for Opus 5.5), long coding and memory, according to the comparison tables reported by VentureBeat. Opus keeps the lead on health, computer use, and a few coding areas.
That split matters commercially. Legal work and long coding are two of the highest-value tasks enterprises pay for, so Wonder Full attacks the most lucrative parts of the market first. The memory advantage is also notable in an era where agents need to hold context across long, multi-step jobs.
The timing looks deliberate. Wonder Full arrives as agentic use cases, extended coding sessions, and memory-heavy workflows become the deciding purchase criteria for enterprise teams. Meta has aimed its first frontier shot squarely at the capabilities buyers are actually testing, rather than at benchmark slivers that rarely surface in real work.
Access and pricing: a break from the Llama era
The biggest strategic shift is that both Muse Wonder models are closed. Meta built its reputation on the open Llama family, and its retreat from that position tells you how seriously the company now treats the frontier as a commercial contest.
Wonder Full is limited to paid Muse Code subscribers. The standard Wonder model is live on Meta’s API at $4 per million input tokens and $20 per million output tokens. That is well below the premium tiers of the top OpenAI and Anthropic models, which puts direct pressure on the pricing strategies of both rivals (VentureBeat).
Why it matters
Meta just turned a two-horse race into a three-horse race. That alone would be news. The unwelcome part for OpenAI and Anthropic is everything surrounding the model: the data centres, the billions of users, and a consumer agent app in Muse that has already found an audience.
Those assets are hard to replicate quickly. A rival can match a model score, but it cannot instantly build billions of distribution points. Meta has spent years assembling exactly that combination, and this release is the first time it has looked genuinely capable of using it at the frontier.
For users, the near-term effect is likely to be cheaper access to top-tier capability. When a third lab competes on price and capability, the winners tend to be the customers. The frontier is no longer a duopoly, and that changes the economics for everyone downstream.
Correction (19 October 2026): An earlier version said Wonder Full ranked second outright, gave Muse Code plan prices of $50 and $200 a month, and quoted Alexandr Wang (described as Scale AI’s) on API pricing. Wonder Full is joint second with Claude Sonnet 5.5; we could not verify the plan prices or the quote, so they have been replaced with Meta’s published API pricing. Wang leads Meta Superintelligence Labs and is Scale AI’s former CEO.
Sources
- Artificial Analysis, Muse Wonder and Muse Wonder Full Intelligence Index results. Reported fact: index scores of 56 and 52
- VentureBeat, Meta’s new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via API. 12 October 2026. Reporting on availability, API pricing and benchmark comparisons
- Wikipedia, Meta Superintelligence Labs. Background
Benchmark scores are third-party (Artificial Analysis) or company-published. The assessment of market impact is the author’s analysis.
How this analysis is researched, sourced and corrected: Editorial standards and corrections. Spotted an error? Contact Philip Hall.
