AI Tech36氪

A Gamble Reshaped by the 'Lobster': Has Meta's Second Comeback Failed?

36氪Oct 14, 20261 min read

AI Summary

Meta has launched Muse Wonder, a new flagship model series built on its personal AI agent app Muse, challenging OpenAI and Anthropic again. While it ranks near the top on benchmarks, developers have criticized its real-world performance, and its high inference costs and pricing are drawing scrutiny.

AI Insights

Meta's Muse Wonder shows strong benchmark results, especially in long-horizon coding and memory tasks, but developer reception and high inference costs (about 183% higher per task than its predecessor) suggest a gap between leaderboard performance and real-world value. The strategy of gating the larger model behind paid tiers while aiming to compress its capabilities into lighter models signals that competitive pricing and efficiency, not just raw scores, will determine whether Meta can regain ground against OpenAI and Anthropic.

Analysis written by AI from the original article. Check it against the source.

Translated text

Meta unveiled a new model series, Muse Wonder (Wonder and Wonder Full), in the early hours of October 13. Currently only Wonder Full is open to paying users. On Meta's personal superintelligence index, Wonder Full scored 59.1, ahead of Claude Opus 5.5 (55.2) and GPT-6 Astra (54.5). However, on Artificial Analysis's composite intelligence index, Claude Opus 5.5 still ranked first with 58 points, while Wonder Full tied with Claude Sonnet 5.5 at 56. The day before the launch, Alexandr Wang said on a podcast that Muse grew out of playing with OpenClaw (the 'lobster') alongside Nat Friedman. Muse Wonder features Active Focus for sustained task persistence and Active Memory for retaining context across conversations, scoring 85.6 on MemoryBench. On the long-horizon coding test SWE-Marathon, Wonder Full scored 57.5%, above Claude Opus 5.5 (46.3%) and GPT-6 Astra (42.5%). Yet developers reported disappointing results, including on a 3D game prototype, and Wang admitted the team had not specifically trained the model for 3D demo games. Pricing is $4 per million input tokens and $20 per million output tokens, matching Claude Opus 5.5. Artificial Analysis estimated the average task cost at $4.52, up 183% from the previous Muse Spark 1.3 and about 39% above GPT-6 Astra. Wonder is therefore initially reserved for paying users, with Meta aiming to compress its capabilities into the lighter Spark form for mass consumers later.

Related News

AI Tech36氪

Anthropic Shocking Study: AI Could Self-Replicate Through TV Boxes

Anthropic's Frontier Red Team published research examining whether AI could replicate, run, and spread across devices autonomously, like a computer virus. The team ran a 26B-parameter model across 15 low-cost Android TV boxes and found that small models can increasingly adapt to unfamiliar hardware. The experiments were conducted in isolated environments, not real-world networks.

AI Tech36氪

Google Awakens to Physics: Gemini 4 Argon Hands-On Tests Pour In

Google's Gemini 4 Argon has launched and reportedly competes closely with GPT-6 Astra and Claude Opus 5.5 on benchmarks, at about one-fifth the price and several times faster. Hands-on tests showed strengths in 3D generation, physics-intuitive visualization, and scientific reasoning, though its Chinese-language narration sounded awkward.

AI Tech虎嗅

Hands-On Test of Gemini 4 Argon: Google's Stumbling Last-Ditch Effort

Google has made Gemini 4 Argon generally available after first announcing it on October 1 for a security program. Following a year of talent departures and delayed model launches, the model is widely seen as a last-ditch effort for Google. Hands-on testing suggests it will struggle to challenge Claude Opus 5.5 or GPT-6 Astra.