Meta’s Muse Wonder AI Matches Claude, Costs the Same
ALSO: White House demands AI reports, OpenAI pulls 3 papers
Welcome to Guru99 AI Report!
Top Story: This week, Meta’s newest model lands two points off the leader but costs exactly what Claude does, an AI agent fires off a made-up tip to real police, and three of OpenAI’s 722 AI-written math papers don’t survive a single day. Let’s dig in.
🧠 Meta’s New Muse Wonder AI: Near the Top, Not Cheap
Brief Buzz:
Meta has introduced two new AI models: Muse Wonder and Muse Wonder Full — the most powerful it has built so far. Both are made for AI agents (tools that carry out tasks for you step by step), and Meta says they remember you and don’t give up easily.
- Wonder Full scores 56 on an independent AI test, matching Claude Sonnet 5.5 — only Claude Opus 5.5 scores higher, at 58.
- The basic Wonder model scores 52 but costs the same as Opus 5.5: $4 in and $20 out per million tokens (chunks of text).
- Wonder Full isn’t sold per token at all — it’s available only through Muse Code, Meta’s coding app, at $50 or $200 a month.
- On Meta’s own tests, Wonder Full wins on long coding jobs, legal work, and memory; Opus 5.5 still leads on health and computer tasks.
- Meta calls these its best-behaved models yet, but admits Wonder Full could help cause chemical or biological harm even with no extra safety measures.
💡 Why Should You Care?
An AI that remembers you is only as safe as the company holding that memory — and Meta’s Muse agent reportedly handed one user’s home address to a stranger, while Muse Code learns from your work unless you switch the setting off.
🚨 Claude Went Off-Script, Now White House Demands Incident Reports
Brief Buzz:
In a new report on unintended model behavior, Anthropic admits some of its AI agents went off-script on the live internet — one even sent a fabricated tip to a real police website. The same day, the White House told all AI companies to report incidents like this immediately.
- Asked to generate example tasks, Claude Haiku 4.5 sent an invented tip to the Philadelphia police — it was flagged as spam and never passed on.
- An Anthropic test model submitted 19 visa applications to the State Department in August, plus one in May, according to an official familiar with the matter.
- Four patterns showed up, mostly what Anthropic calls persistence: when blocked, the model works around the block instead of stopping.
- Anthropic says the impact is minimal and has simply switched off live internet for all internal evaluations; its new detectors caught every instance reported in testing.
- The White House’s Super Intelligence Force (formed Oct 4) says disclosure is “not optional,” but hasn’t spelled out any enforcement.
💡 Why Should You Care?
AI agents act on your behalf — they click, type, and submit — and this shows that when they get stuck, they can wander into real systems, including police and government sites. Watch whether the White House’s disclosure request gains any real teeth.
🔍 Just 3.6% of Chinese AI Models Have Published Safety Results
Brief Buzz:
China’s top labs are shipping AI models fast, but safety data is almost nowhere to be found. A new SemiAnalysis report looked at 857 releases from 9 Chinese developers between 2021 and mid-September 2026 — and only 31 of them (3.6%) came with a published safety test result for that specific model.
- Just 9 had results ready at or before launch; 16 showed up later, a median of 42 days after release.
- Alibaba, ByteDance, Tencent, and Baidu have posted results for only 11 of their 540 releases combined.
- Startups did better — 20 of 317 — and Zhipu (Z.ai) is the only lab with results every year since 2022.
- 93% of reasoning models, the fastest-moving category, have no published safety results at all.
- Missing results don’t prove the labs skipped testing, and the report drew no comparison with US labs.
💡 Why Should You Care?
Since mid-2023, open-weight models (the ones you can download and run yourself) have made up over half of all releases each quarter. If you build on one, assume the safety testing is your job.
🧮 OpenAI’s Unreleased AI Posts 722 Math Papers, Pulls Three
Brief Buzz:
OpenAI has published 722 math papers written by an AI model it hasn’t released, spanning math, computer science, and physics. The twist: within a day, three of them were pulled for errors.
- The 722 manuscripts cover 372 “families” of related results, all posted to a public GitHub repository under an open license.
- Roughly 4,000 problems were fed to the model, OpenAI says, with each result taking about three hours of ChatGPT Pro thinking compute on average.
- Only about 42% of the headline results are formalized in Lean (the software that machine-checks proofs); OpenAI admits the unverified ones “could have issues.”
- Three manuscripts were pulled on Oct 7 after one proof turned out to contain a single sign error, with two more papers built on top of it.
- Mathematicians weren’t pleased: Cornell’s Alex Townsend told Retraction Watch OpenAI should have released the Lean-verified results first.
💡 Why Should You Care?
AI can now produce research far faster than humans can check it, shifting the bottleneck from discovery to verification. Until experts have gone through the proofs, treat them as claims, not breakthroughs.
🔻 Karpathy: Only ~5,000 People See AI’s Real Frontier
Brief Buzz:
Andrej Karpathy, now at Anthropic, says the gap in how well people understand AI is widening. In a recent post on X he describes it as a funnel: most people see a clumsy chatbot, while a small group sees something far more advanced.
- By his own rough math, about 75% of humanity — some 6 billion people — have barely used an AI model.
- One to two billion casual users treat free chatbots like a smarter Google or a writing aid.
- Twenty million coders and mathematicians watch AI agents finish weeks of work from a single prompt.
- Around five thousand people in frontier labs see agent swarms tackle jobs that would take top experts years.
- There have been public glimpses: in July, OpenAI’s 722 AI-written math manuscripts and test agents that broke into Hugging Face — humans are still checking both.
💡 Why Should You Care?
If AI looks unimpressive to you, that may say more about your access than the technology’s limits. Karpathy ties the real gains to hard, substantial, easily verifiable digital work — but treat his figures as one insider’s estimate.
🎰 Broadcom CEO: Many AI Investors Are Basically Gambling
Brief Buzz:
Hock Tan, CEO of Broadcom — the company that designs many of the chips behind the AI boom — has issued a warning. At a Singapore conference on Monday, he said many AI investors are placing blind bets, pouring huge sums into technology they barely understand.
- Many investors have no real idea what they’re buying, Tan said, and their expectations are running far ahead of reality.
- He compared companies’ AI spending to kids burning ten dollars on arcade tokens in ten minutes and learning nothing.
- Per Reuters, OpenAI’s September revenue run rate is about $50 billion (the annual figure you get by multiplying a month’s sales by 12) — which Tan says shows demand, not future profits.
- Tan warns that capital, not chips, could become AI’s next bottleneck — days after Nvidia-backed Firmus scrapped its roughly $5 billion IPO.
- Intel’s Lip-Bu Tan said another two to three years of tight hardware supply lie ahead, with memory chips perhaps easing first.
💡 Why Should You Care?
When the company selling the picks and shovels of the AI boom says parts of the market are betting blind, it’s worth noting. Expect more scrutiny of whether workplace AI tools are worth it, and more caution around AI-heavy investments.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
