Tokens by Michael Brenndoerfer, Week #21

Tokens

Claude Mythos 5.2 tackles black holes and math proofs

Today's issue: Anthropic unveils Claude Fable and Mythos 5.2, Meta discounts agent tokens, and DeepMind proves Signal encryption is solid.

Welcome to #21 of your weekly AI newsletter.

Don't miss the next one 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

No spam. Unsubscribe anytime.


Smiling cartoon computer

Main Story

Train Hitting a School Bus meme depicting Claude Mythos instantly crushing centuries of mathematical effort in thirty hours.
Fermat spent decades working on his theorem; Claude Mythos knocked it out before the weekend was over.

Anthropic drops Claude Fable 5.2 and locks Mythos behind safety glass

Anthropic rolled out Claude Fable 5.2 and Claude Mythos 5.2, pushing its reasoning, coding, and vision stack into territory previously reserved for graduate seminars and research labs. Both models share the same architecture, generate text up to 30 percent faster than their predecessors, and communicate with noticeably cleaner phrasing. Only one of them is actually reaching the public.

During prerelease testing, the dual models reportedly solved two open problems in black hole physics, designed three working electronic hardware devices from scratch, and completed a formal proof of Fermat's Last Theorem in 30 hours. For comparison, mathematicians spent centuries batting that problem around, and Andrew Wiles spent years in an attic to deliver the human proof in the 1990s.

The split release shows how frontier labs now handle extreme technical capabilities. Claude Fable 5.2 gets broad availability, bundled with strict biological and cybersecurity safeguards. Claude Mythos 5.2 stays behind Anthropic's velvet rope, reserved exclusively for vetted partners in its Life Sciences and Cyber Verification Programs.

  • Rigorous proofs: Formalized Fermat's Last Theorem in 30 hours and tackled two open black hole physics questions during testing.
  • Physical engineering: Successfully designed three working electronic hardware devices completely from scratch.
  • Tiered availability: Fable 5.2 ships with general safeguards, while the Mythos tier stays restricted to vetted research programs.

Bottom Line: As frontier models graduate from writing boilerplate scripts to solving foundational scientific equations, open-door releases for flagship weights are effectively dead. When a model can draft working electronics and crack theoretical physics out of the box, gatekeeping becomes standard operating procedure.

Source: www.anthropic.com

Don't miss the next one 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

No spam. Unsubscribe anytime.


Smiling cartoon star

The Feed

  1. DeepMind put the Signal protocol through its AlphaProof Code framework over four weeks, attempting to break its cryptographic foundations. The agentic system mathematically verified the underlying math, finding zero cracks.

  2. Meta launched Muse Ember, a compact reasoning model priced at ten cents per million input tokens. Built for repetitive agent chores like tool calling and routing, it matches Claude Haiku 5.5 performance at one-third the price.

  3. NVIDIA unveiled GR00T 2, an open world action model trained on human video and gripper demos. It hits a 95% success rate across 50 manipulation tasks with five fine-tuning examples, and 40% in zero-shot robot transfer.

  4. Google dropped TPU 8tx and TPU 8i Max, expanding its eighth-gen silicon lineup to four systems in one year. The hardware focuses on securing agentic workflows and feeding Gemini models over huge context windows.

  5. Australia is mandating safety audits for frontier AI developers under the Australian AI Standards. A dedicated safety regulator and the Australian AI Safety Institute will oversee compliance and enforce model evaluations.

  6. Tencent released Tencent Hy4, an open-source model adding visual math reasoning and fine-grained image processing. The release includes full model weights, cybersecurity safeguards, and native GUI computer-use agents.

  7. Researchers built AutoDNA, an autonomous lab platform coordinating over 20 scientific instruments across 11,000 steps. Rather than treating AI as an add-on assistant, it co-designs experiments and runs nucleic acid engineering.

  8. Google reached general availability for Cloud Pathways on Google Kubernetes Engine. The distributed runtime coordinates large accelerator fleets, logging 130 million Cloud TPU chip-hours across GKE over the past quarter.

  9. Baidu dropped ERNIE-5.5-Flash-Base directly onto Hugging Face. The 280-billion-parameter open-weight base model activates 14 billion parameters per token across 42 mixture-of-experts layers without chat tuning.

  10. Ant Lingbo introduced LingBot-Touch, a tactile foundation model starting at 5.3M parameters. Trained on 4,983 hours of touch data, it bumped robotic zipper-closing success on the Sharpa Wave hand from 39% up to 59%.

  11. Researchers published Axion 1.0, an open foundation model built specifically for AI agents. The project open-sources weights, training code, and filtering data while ditching complex steps like DPO and on-policy distillation.

  12. NYU Langone deployed Lang1, a clinical prediction engine trained on records from 10 million patients. Despite its compact footprint, it beat models like DeepSeek R1 by up to 24 percent when forecasting patient decline.

  13. Users are uploading empty house photos and letting ChatGPT furnish them based on stored memories. An analysis of over 20,000 interactions found the viral trend creates an accurate algorithmic self-portrait in seconds.

  14. Adventists AI open-sourced Sophon-Realtime v1, an Apache 2.0 voice family running on Qwen3-VL-32B-Instruct. The full-duplex models detect conversation pauses and insert natural audio cues like laughter and sighs.

  15. A new gait foundation model analyzed 3D skeletal movement data from 3,328 adults using depth-sensing video. The movement embeddings reliably predicted 1,887 phenotypic traits across nearly all human organ systems.

See you next week!

Michael

Smiling cartoon mobile phone

Do you want to be mentioned in the newsletter, or do you have feedback?

Get in Touch
Smiling cartoon computer

Get Tokens in your inbox 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

Join 5,000+ readers · No spam, unsubscribe anytime.