How Muse Wonder Shifts the AI Landscape
Meta's Muse Wonder models put it back at the AI frontier on benchmarks, though real-world results are mixed, and Meta's fast pace and 'normal technology' stance set it apart from rival labs.
Watch on YouTube ↗Charlie's summary
- ▶ 0:33
Speed is becoming a new axis of competition among frontier labs, even though fast modes are still a novelty.
Anthropic now serves Opus 5.5 at a verified 285 tokens per second in fast mode, and OpenAI's docs suggest a theoretical maximum of 280. OpenAI charges 50% more for its premium speed. Both are expensive and niche for now. The host thinks big shifts in latency will change how we interact with models in ways “we'll quite know until we get there.”
- ▶ 4:08
Amazon's Decart bid suggests it is betting that hardware efficiency is the only benchmark that will matter.
Amazon is reportedly negotiating to buy Decart at a $7 billion valuation, a 75% premium on its last mark. Decart's optimization system claims inference up to eight times as fast at one-eighth the cost, and it is hardware agnostic. The host reads this as Amazon preparing for a world focused on serving tokens cheaply, though it is unclear whether the technology applies to general LLMs.
- ▶ 6:13
Everyone now competes with everyone: chips, models, clouds and apps are converging, so the question is what the durable constraint is.
Nvidia is reportedly eyeing Reflection AI after buying or acqui-hiring other open-model players, and Microsoft unveiled its Maia 300 chip, which it says delivers 40% more tokens per watt than leading GPU hardware. Investor Chimath's view, as quoted: silicon companies move up the stack as application companies move down. The open question is whether distribution, data or physical constraints like power will anchor long-term margins.
- ▶ 9:46
On benchmarks, Muse Wonder Full puts Meta back at the frontier for the first time in years, and just weeks after Muse Spark.
Wonder Full scored 79% on Deep Suite 1.1, almost five points ahead of GPT6 Astra and Opus 5.5. Artificial Analysis gave it 56 on its intelligence index, tied for second and two points behind Opus 5.5. It comes only six weeks after Muse Spark 1.3. The cost is high, at $4.52 per task, almost 3x Spark 1.3. Shopify's Toby wrote: “Meta is back.”
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.