Meta just caught up with Anthropic

Video thumbnail: Meta just caught up with Anthropic
Oct 13, 202638m 2s video lengthTheo - t3․gg

The Signal

Meta has released a new model family, Muse Wonder and Muse Wonderful Full, positioning them as a third serious contender alongside OpenAI and Anthropic. The speaker, a developer who tested the models, asserts they are uniquely effective at workflow orchestration, bug finding, and generating readable reports, though they explicitly perform poorly on 3D and UI tasks. The central tension is between the model’s surprisingly effective operational behavior and its product fragmentation; the more capable 'Wonderful Full' model remains behind a subscription wall with opaque pricing, while only the weaker 'Wonder' model is available via public API.

The Case

Model Performance and Workflow

  • Muse’s standout strength is its ability to act as a 'commander'—orchestrating sub-agents and managing large-scale coding tasks—without the catastrophic mistakes that frequently plague other models.16:55
  • In side-by-side evaluations, Claude/Opus remains superior for deep reasoning and complex correctness, while Muse is often faster and produces more readable, actionable audit reports.14:19
  • The model is intentionally weak at 3D rendering and front-end UI generation, a limitation Meta acknowledges and the speaker attributes to specialized post-training environments rather than a general intelligence failure.6:36
  • The speaker values Muse’s 'low-dumbness' floor, observing fewer operational blunders—like file deletion errors—compared to models like Astra or Grok.18:57

Product and Pricing Transparency

  • The model family is split into two distinct tiers: Muse Wonder, which is API-accessible, and Muse Wonderful Full, which requires a subscription, creating an evaluation gap because most benchmark wins were achieved using the version users cannot access.3:04
  • Pricing for the 'Wonderful Full' model is unpublished, which prevents users from conducting a true price-to-performance analysis despite its strong performance on benchmarks like Artificial Analysis.3:51
  • The speaker recommends against replacing existing Cloud Code subscriptions with Muse until pricing and usage limits for the full version are transparent.36:21

Practical Implementation

  • For high-risk or deep-reasoning tasks, the speaker continues to rely on Opus, using Muse primarily as an efficient reviewer or diagnostic assistant that can identify existing fixes like one-line PR changes.36:00
  • Planning reports generated by Muse—such as those prioritizing T3 Code development—favored concrete 'kill or freeze' actions over broad strategy, which the speaker found highly usable.33:16

The 1 Minute Signal Take

Muse is a genuinely competitive third player that succeeds by prioritizing workflow reliability and readability over raw benchmark supremacy. While it is not a universal replacement for the current leaders, its ability to orchestrate tasks effectively makes it a valuable addition to a developer's multi-model toolkit, provided you treat it as a specialized diagnostic tool rather than a replacement for high-stakes reasoning.

Pro Analysis

Why It Matters

Meta’s entry proves that frontier-level intelligence is no longer restricted to just two companies, which typically forces aggressive pricing and quality improvements across the entire ecosystem. This commoditization of "orchestration-ready" AI changes the calculus for software engineering workflows from single-model dependency to multi-model diversification.

Strategic Implications

Businesses now face a "model selection tax," where different phases of the software development lifecycle (SDLC) require different engines. The fact that Muse is excellent at command and control but poor at front-end visual tasks suggests that future AI-integrated IDEs will likely feature dynamic model-routing as a core product feature rather than an afterthought.

Evidence & Hype Audit

This content is a high-signal report based on personal hands-on usage rather than marketing collateral. While the speaker admits to a preference for "novelty" and a distaste for specific competitors, the use of side-by-side performance metrics—and the explicit admission of Muse's UI/3D failures—provides a necessary grounding that validates the speaker’s objectivity.

Counterarguments

Critics might argue that Muse's "readability" is merely a stylistic preference or a form of lossy summarization that hides edge-case complexity. Furthermore, the reliance on an external company for 3D training environments, as speculated, suggests that Muse might be benefiting from borrowed capability rather than architectural superiority.

Who Should Care

  • Engineering Leads: Looking to optimize agentic workflows for speed and review quality.
  • Product Managers: Needing to know if their current AI-integrated tools are using the "full" version of a model or a distilled API version.
  • AI Platform Engineers: Monitoring the shift toward multi-model orchestration.

What to Do Next

  • Conduct an A/B test of Muse versus your current primary model on a standard repository audit task.
  • Review current spending to determine if a multi-model subscription strategy would lower costs compared to raw token usage.
  • Create a "kill list" of tasks—like 3D/UI—that are explicitly banned for your team's AI automation pipeline to prevent output "slop."
Time saved:34m 15s

Share this

Tags

Written by: 1 Minute Signal Editorial Team