Why Muse Wonder Is Dividing AI Users

The episode opens with a review of faster model-serving options. Anthropic is reportedly serving Opus 5.5 at a verified maximum of 285 tokens per second through a new fast mode, while OpenAI’s documentation suggests a theoretical maximum of 280 tokens per second for GPT-6.1 Soul’s fast mode. OpenAI charges 50% more for premium speed, and Anthropic’s faster usage is not included in ordinary subscriptions. Chubby estimates that Anthropic can provide 50% more tokens per dollar for users trying to process large volumes quickly. The host says these modes are still a novelty for most users and are expensive enough to have niche applications, but he expects latency to change how people interact with models as fast service becomes standard. He then reports sightings of a new Anthropic model receiving queries that previously went to Opus 5.5, including code-animation and Fable prompts. Chubby predicts a release soon, possibly timed ahead of Anthropic’s IPO so the market can incorporate the model into its valuation. The episode frames the release as a test of whether major labs will pace the frontier through mid-sized models or accelerate again with larger systems. An AI dinner attendee told the host that Fable 5.5 would be “stunning,” but that OpenAI would beat it in December with Bell.

OpenAI is reportedly discussing a Sora revival with a stealth startup founded by Bill Peoples, the former OpenAI researcher who led the Sora project. Sora briefly reached number one on the App Store in late 2025, generating a wave of TikTok-style videos and cat memes, but OpenAI mothballed it in April when it prioritized enterprise uses such as coding and ended what it called sidequests. The potential partner also includes former Disney chairman Jeffrey Katzenberg and former Dropbox CFO Sujay Joshua. The deal may not provide full access to Sora’s weights, but the startup would reportedly receive enough access to customize the model heavily. The host also discusses Amazon’s negotiations to acquire Decart AI at a reported $7 billion valuation. Decart develops world models for live video, including character transformations and generative games, but Amazon’s interest appears to center on its Decart Optimization System, or DOS. After early access to Amazon’s Trainium 3 chips, Decart claimed DOS could make inference up to eight times faster at one-eighth the cost, and that it was hardware-agnostic. Amazon reportedly considered a $6 billion deal last month before walking away during due diligence, while Decart’s last valuation was $4 billion. The host says the acquisition would signal that Amazon is betting on inference efficiency rather than trying to win the model race directly.

Nvidia is reportedly considering a further investment, cash-and-compute arrangement, or full acquihire of Reflection AI after the open-model developer released Beam. Reflection was last valued at $25 billion, and Nvidia already owns an $800 million stake, so a full transaction could exceed Nvidia’s $20 billion acquisition of Groq and become its largest acquisition. Beam was intended to rival leading Chinese open-weight models; although it did not match them on every benchmark, it was considered a meaningful model release. The host interprets Nvidia’s activity as evidence that Jensen Huang wants a strong Western open-model stack, citing Nvidia’s investments or acquisitions involving Hugging Face, Poolside, Essential AI, and now Reflection. One counterargument says Nvidia is more likely trying to improve NemoTron than become a direct rival to OpenAI and Anthropic; competing as a model lab could encourage those companies to move away from Nvidia hardware. The episode then turns to Microsoft’s Maia 300 custom AI chip. CEO Satya Nadella called it Microsoft’s biggest custom-silicon bet and claimed it would deliver 40% more tokens per watt than leading GPU hardware. Microsoft plans multi-gigawatt deployments beginning the following year, making Maia the primary inference system for its own products and services. The host connects this to the broader idea that every company may soon offer models, silicon, cloud infrastructure, and applications, leaving distribution, data, power, or hardware efficiency as possible durable constraints.

The main segment examines Meta’s Muse Wonder family, released after the earlier Muse Spark models. Wonder comes in a faster, slimmed-down version and Wonder Full, the larger model available through an API. On DeepSweep 1.1, Wonder Full scored 79%, nearly five points ahead of GPT-6 Astra and Opus 5.5. On Terminal Bench 4.0 it was only two points behind those rivals, and on Meta’s personal superintelligence index it led Astra and Opus by about four points. It also reportedly led on Harvey Lab’s legal benchmark, Val’s finance-agent benchmark, and Automation Bench, while trailing the top systems by a few points on OSWorld computer-use testing. Artificial Analysis gave Wonder Full a 56 on its Intelligence Index, tied for second with Sonnet 5.5 and two points behind Opus 5.5. The slimmer Wonder scored 52, tying GPT-6.1 Soul and the unreleased Gemini 4 Argon in seventh place. Artificial Analysis called the release a material gain over Muse Spark 1.3 only six weeks later, particularly in coding and professional-work tasks. The price was a significant drawback: Wonder Full averaged $4.52 per task, almost three times the cost of Spark 1.3, roughly one dollar more than GPT-6 Astra or Grok 4.7, and about $1.50 less than Opus 5.5. Based on benchmarks alone, the host says Meta is back in the frontier conversation.

The initial enthusiasm was challenged by users testing Wonder on practical tasks. Max Blade compared Wonder and Opus 5.5 on building a AAA-level lawnmower simulator and found Meta’s model far behind on 3D game work and printable CAD designs. The test consumed a $50 spending plan plus $4.11 in equivalent API tokens, forcing an upgrade to a $200 plan. Meta Chief AI Officer Alexander Wang acknowledged that the team had not explicitly trained on 3D demo games and asked users to evaluate Wonder on actual coding instead. The host agrees that optimizing for visually impressive 3D demos may be harmful because social-media reactions pressure labs to target a narrow use case. Other reports were mixed or negative. Mehumahen said Wonder Full was nowhere near Opus 5.5 and possibly behind GPT-6.1 Soul on real programming tasks. Dan Shipper reported that it scored 52% on a benchmark requiring a model to rewrite a database from scratch, compared with 71% for GPT Soul. Riley Brown used it for five hours and returned to Codex and Claude Code. Several users thought Wonder Full was strong but the slimmer Wonder was poor, creating confusion because both products share the same name. Usage limits were another problem: one user said Wonder consumed 13% of a five-hour allowance in under ten minutes, while Wang suggested a bug may have miscounted usage. The positive reports were specific rather than universal. Theo placed Wonder Full in his top tier after it fixed a bug, while GPT-6.1 Soul crashed during a tool interaction and produced an overlong explanation. Charlie Holtz said it found a real security vulnerability other agents missed and was strong at design, visualization, sub-agents, and CLI details. Another tester called it technically competitive, creative, warm, and unusually aligned with the user. Clarevo found it weak at front-end design, SVGs, video, and creative work, but first among eight frontier models on her agentic-task tests and competitive on coding. The episode’s central conclusion is that Wonder’s benchmark standing does not predict a consistent experience, especially when comparing the two variants.

The host places Wonder in Meta’s larger consumer-agent strategy. Muse had been notable because most users did not care which model powered it, but users increasingly encountered limits: one person said Muse hallucinated during a multi-step task and compounded its own errors, while others already used several assistants—Instinct, Muse, Grokbot, and Dots—for different purposes. The host argues that personal assistants differ from ordinary LLM or coding-agent tools because their value rises with access to personal context and systems. Since most people will struggle to give even one assistant complete access, using several assistants may impose too much context-switching. Meta appears to be addressing failures in long-running tasks through two features associated with Wonder. Active memory tracks what matters across conversations, while active focus keeps working when a task becomes difficult, tries alternative approaches, and recovers from errors. These features target reported problems such as Muse giving an incorrect reminder about a meeting that had occurred the previous day or getting stuck in a Hilton hotel-booking loop. The host also notes that Meta repeatedly described Wonder as state of the art for both personal and professional agents. He interprets this less as an immediate attempt to sell enterprise coding tools against products like Cloud Code and more as recognition that a personal assistant must eventually handle users’ professional tasks. Small-business owners may be especially important because they already use Facebook and Instagram extensively, and Citrini Research says Meta has a credible balance of capability and distribution for that market.

The episode argues that the most important fact may be the speed of Meta’s progress rather than Wonder’s exact endpoint. Muse launched only 34 days before Wonder, and commentators including Ethan Mollick described the rapid improvement as more impressive than a limited test of the model itself. You Chen Jin credited Meta’s decision to spend heavily on top talent, assemble a strong team, and give it an ambitious goal and sufficient resources. Meta’s stock rose 4% after the news. Wonder also reinforces a philosophical split over frontier development. Mark Zuckerberg did not sign the letter calling for a pause or slower frontier progress; instead, his essay “The Future is for everyone” argued that broad access should be central even if openness creates additional risk. The host says Wonder is not itself pushing the frontier and appears to have strong cyber guardrails, but Zuckerberg told the Financial Times that Meta had reached the frontier and was already working on the next step. He predicted model development would probably move faster and said he was closer to viewing AI as a technology that can be engineered with normal guardrails than as an autonomous, anthropomorphic force. The host summarizes the difference as follows: OpenAI and Anthropic are either building or inevitably moving toward machine gods that inspire awe, while Meta is treating AI as a normal technology. At the start of 2026, the United States seemed to have a frontier group centered on OpenAI, Anthropic, and Google, later narrowing to two as coding agents dominated attention. With Meta’s return—and potentially Google’s unreleased Gemini 4 Argon—the host sees roughly five frontier labs competing again, even if another wave of stronger models could quickly restore a duopoly.