Meta's Muse is a real contender, but you can't price the good

The surprise. Meta's latest model release, Muse, genuinely competes with top-tier models from Anthropic and OpenAI—an outcome the reviewer did not expect. The catch: only the cheaper, less capable 'Muse Wonder' ($4/M in, $20/M out) is available via API. The truly competitive 'Muse Wonderful Full' is subscription-only with no published price, making cost comparisons impossible.

The split makes it messy. The two-model release creates confusion. The smaller Wonder model performs poorly on real-world audits, scoring nearly half on code analysis tasks compared to Anthropic's Opus. The full version matches or beats OpenAI's Astra and Fable on benchmarks but only exists within Meta's subscription service. This split means the public API version underrepresents Meta's actual capability.

What stands out: low dumb-factor. The reviewer emphasizes that Muse makes fewer 'dumb' mistakes than competitors. It rarely deletes wrong files or enters broken states. Its responses are more readable and concise. In one example, Muse found a relevant PR fix that both Opus and another model missed. The trade-off is a lower intelligence ceiling—roughly 8-9/10 vs Opus at 10/10.

The orchestration edge. Muse handles model delegation unusually well. It writes good prompts for sub-agents, coordinates work across multiple models, and manages large parallel tasks. This is rare for a model outside Anthropic or OpenAI. In tests where Muse and Opus collaborated, Muse's plans were often preferred as the base, with Opus contributing refinements.

The hype problem. The evidence is almost entirely anecdotal. The reviewer admits they enjoy the 'new flavor' of responses and may be biased. There are no controlled benchmarks for the full version, no pricing transparency, and the reviewer is using a free API key rather than a paid subscription. The 3D and UI capabilities are notably weak, which Meta itself acknowledges.

What this means. Meta is back in the frontier race, but the opaque pricing model and split release strategy are anti-consumer. Engineering teams evaluating Muse should ignore the public API version entirely and treat the subscription-only model as unmeasurable. The real signal is that three credible model families now exist, which should drive down costs and increase innovation across the board.