$1.8B AI Biology Investment to Build a Virtual Cell

Caught My Eye…

1) OpenAI's Math Breakthroughs Raise a Crypto Security Question

On October 6, OpenAI revealed over 700 mathematical manuscripts, produced by an unreleased internal AI model. OpenAI says about 42% of the headline results have proofs formalized in Lean, a tool that verifies mathematical reasoning.

Computer scientist Scott Aaronson noted that cryptography is conspicuously missing, and wrote that his sources say AI companies have begun discreetly investigating whether their latest models can break important cryptographic protocols.

The release has sparked debate over whether AI could find a shortcut through the mathematics that protects cryptocurrency wallets. Bitcoin and Ethereum use digital signatures to authorize transactions. A secret key produces the signature, and a public key lets everyone else check it. Their security depends on the difficulty of recovering the secret key based on the public key.

A powerful enough quantum computer could do that using a known algorithm, but the new concern is that AI could discover a shortcut that works on ordinary computers.

Ethereum Foundation researcher Justin Drake urged planning for a controlled move into addresses that keep public keys hidden until use. His worst-case scenario was a break within months, and he warned against rushing, as panic causes more issues.

Ethereum co-founder Vitalik Buterin also recommended keeping funds in addresses that have not yet been used to make a transaction. Coinbase's head of cryptography, Yehuda Lindell, responded by saying that there is currently no evidence of anything major happening, so there is no need to rush. Instead, he recommended that blockchains add hybrid schemes using different math methods along their post-quantum migrations to increase security.

2) The Hidden Cost of Cheap AI Models

On October 7, Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens, the same list price as OpenAI's GPT-6 Luna for prompts under 100,000 tokens. This price range also matches frontier Chinese models while delivering similar performance.

A study by researchers at Stanford, UC Berkeley, Carnegie Mellon and Microsoft Research shows how actual final costs drift from list prices. They ran eight reasoning models, which work through a problem step by step before answering, on 12 kinds of tasks. In 32% of 336 head-to-head comparisons, the model with the lower list price cost up to 28 times more in total. At the study's May 1 prices, Google's Gemini 3 Flash listed 80% cheaper than OpenAI's GPT-5.4, yet running every test cost $705 on Gemini and $509 on GPT-5.4.

The reasoning steps, called thinking tokens, are billed at the higher price for text the model writes. On one-question tasks where the cheaper-listed model cost more overall, thinking tokens explained more than 95% of the gap. On one test set, Gemini 3 Flash produced more than 71 million thinking tokens, compared with GPT-5.4's 1.8 million.

The buyer sees the price per token, but the model decides how many tokens a job takes, within an effort setting the buyer picks. Haiku 5.5 is Anthropic's first small model with an adjustable effort setting, and Anthropic says its updated tokenizer, the software that splits text into tokens, uses slightly more tokens per task. The study ran each model at a single reasoning setting and did not weigh cost against answer quality, and it predates this week's releases.

To learn more about what results companies are seeing from AI spend, read our recent Deep Dive on AI ROI here.

3) $1.8B AI Biology Investment to Build a Virtual Cell

On October 7, Biohub, the nonprofit research institute founded by Mark Zuckerberg and Priscilla Chan, expanded its effort to build a virtual cell. The idea is to create an AI-powered simulation of a living cell that can predict how it responds to drugs, genetic changes, and other conditions. Scientists could then test ideas on the virtual cell before deciding which experiments to run in a real lab.

To make this possible, Biohub and its partners plan to use robotic labs to test how living cells respond to thousands of different conditions. The results will then train AI models that can predict how cells behave.

Biohub originally committed $500 million to this effort in April. This week, it announced over $800 million in additional commitments, including more than $500 million from the Department of Energy over five years and a combined $300 million from Meta, Google DeepMind, and Isomorphic Labs, Alphabet's AI drug development company. Isomorphic is also in early talks to raise at an over $40 billion valuation.

The National Institutes of Health is also contributing existing datasets developed through more than $500 million in past federal spending. Together, Biohub values the initiative's funding and scientific resources at $1.8 billion.

As an incentive to help fund this endeavor, commercial partners get one year of exclusive access to the data they develop before it is released to the public.

AI can already model protein shapes, but a whole living cell is far more complicated, and the data a model would learn from is currently lacking. Biohub expects the first dataset will be ready in about a year.