Meta just caught up with Anthropic

Summary

The presenter reviews Meta’s new Muse Wonder models and argues that the larger Muse Wonderful Full is Meta’s first model that feels competitive with leading “frontier” models from Anthropic and OpenAI. He says it is unusually pleasant for coding and long-running agent workflows, though it has important weaknesses and its value is hard to judge without pricing for the full version.

Models, access, and pricing

• Meta released two versions: Muse Wonder and Muse Wonderful Full. The reviewer says they differ substantially in capability, despite their similar names.
• In his tests, the regular Wonder ran at about 120 tokens per second, while Wonderful Full ran at about 50 tokens per second.
• Only the regular Wonder is available through an API with published pricing. The presenter quotes $4 per million input tokens, $20 per million output tokens, and $0.20 per million cached input tokens.
• Wonderful Full is available through a Muse Code subscription, but Meta had not published its price. This makes its price-to-performance difficult to assess, and the presenter speculates it may be expensive.
• He also did not test the $50 or $200 subscription plans, so he could not evaluate their usage limits.

Strengths highlighted

• Distinctive, readable coding assistance: The presenter found the full model’s answers concise, well organized, and easier to follow than those from some competing models. He emphasizes that its style feels different rather than like another imitation of Claude.
• Strong agent orchestration: In his tests, Muse Wonderful Full was effective at delegating research to sub-agents, giving them clear instructions, and coordinating multiple tasks. He considers this a notable strength for a model outside the established leading labs.
• Useful for practical coding work: Across an ongoing compiler project and reviews of his codebases, Muse handled fixes, standard-library work, fuzzing, and code review effectively. He describes it as fast and capable enough for much of his day-to-day work, while reserving harder design decisions for Opus.
• A high “floor” in his experience: The presenter says the model made fewer obviously foolish mistakes than several alternatives he tested. He characterizes it as having somewhat lower peak capability than Opus, but an especially low tendency toward careless errors.
• Efficient runs: In his informal codebase audits, Wonderful Full often completed much faster than Opus or Sonnet and used fewer output tokens than several comparison models. He cautions that these results do not establish cost efficiency because the full model’s price is unknown.

Weaknesses and caveats

• Poor UI generation: The presenter describes the model’s user-interface output as especially weak and advises against choosing it for polished frontend work.
• Weak 3D work: Its Blender and 3D-game results were poor compared with stronger competitors. The presenter notes that Meta itself lists visual and spatial reasoning, particularly 3D tasks, as a limitation.
• The regular model may not represent the full model: In one codebase-analysis comparison, regular Muse Wonder scored far below Wonderful Full. The presenter considers the standard API model much less compelling, especially at its listed price.
• Benchmarks are not the whole story: The reviewer discusses benchmark results but also stresses that many of his most favorable examples involve readability, judgment, and finding useful existing fixes—qualities that are difficult to capture in standard benchmarks. His codebase tests are informal and should not be treated as definitive rankings.
• Not a replacement recommendation: He does not advise cancelling a Claude/Cloud Code subscription. His positive assessment applies mainly to Wonderful Full, whose cost and subscription limits were still unknown.

Tests and comparisons discussed

• Meta’s published benchmark claims for Wonderful Full, including results on Terminal Bench, DeepSuite, and SW Marathon.
• An Artificial Analysis ranking that the presenter says appeared to place a Wonderful Full run third overall.
• Skatebench, where he reports a 95% result for Wonderful Full—close to Gemini models’ unusually high scores on that test.
• Informal comparisons of models auditing large codebases, reviewing pull requests, and proposing engineering priorities.
• Side-by-side coding-agent examples involving Opus, Sonnet, Astra, Fable, Soul, and Muse. These comparisons are based on the presenter’s own projects and methods, not a formal standardized evaluation.

Overall assessment

The presenter’s conclusion is that Meta has produced a genuinely competitive model with a distinctive interaction style, strong agent coordination, and useful coding performance. He expects to keep using Wonderful Full, especially for routine engineering tasks and reviews, while bringing in Opus for more complex or high-stakes work. He says its arrival matters even beyond its individual performance because another serious competitor could increase pressure on other AI labs to improve model quality and pricing.

Sponsor segment

The video also promotes Browserbase as infrastructure for web-enabled agents. The presenter highlights browser navigation, a credential system that integrates with 1Password, and separate search and page-fetch APIs that can return web content in formats such as HTML, JSON, or Markdown.

Main speaker and sources

• Main speaker: Theo, the video’s presenter and reviewer.
• Sources discussed: Meta’s model announcement and benchmark claims; the presenter’s hands-on coding and agent tests; Artificial Analysis and other named benchmarks; comparisons with models from Anthropic, OpenAI, Google, and other labs.