Tuesday, October 27, 2026HotTea archive editionVerified 12:33 AM PDT

The lead

Anthropic files for an IPO with $16.3 billion in half-year revenue

The filing reports a $693 million operating loss. Changes in the value of financing instruments account for most of the much larger net loss.

Listen to this edition

Prefer audio? The briefing has chapters and a full transcript.

The briefing

The rest of the morning

4 more stories

02

OpenAI cuts the cost of its Terra model and targets routine computer work

OpenAI released GPT-6.1 Terra on October 26 for ChatGPT Work, Codex, and its developer platform. The company says its standard developer prices are half the price of GPT-5.6 Terra and GPT-6.1 Sol. OpenAI reports a 15 percentage point gain over GPT-5.6 Terra at medium effort on Agents' Last Exam. It says that result cost about one-third as much.

These are OpenAI’s measurements, not an independent test of an employer’s workflow. A lower model bill does not include the cost of checking work, fixing mistakes, or controlling access. OpenAI says the model reaches its Critical cybersecurity capability threshold; that rating is not evidence that every deployment is safe.

OpenAI ↗The Next Web ↗
Read article →
03

OpenAI says its AI refuted a major computing conjecture. Other researchers now need to check it

OpenAI announced AI-produced results in mathematics, computer science, and physics on October 26. It says one result refutes the Exponential Time Hypothesis, an assumption used to set conditional limits on algorithms. OpenAI says it formalized the result in Lean. It names Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams as leaders of the group that will examine and explain the work.

This is OpenAI's claim, not an independent confirmation by HotTea. The company says the computer-science output is public. It says the physics group has chosen to withhold its AI-produced artifacts while preparing a write-up. Those release choices allow different levels of outside scrutiny. A claimed ETH refutation does not by itself establish a solution to P versus NP or a practical speedup.

OpenAI ↗
Read article →
04

UF finds that more patient data helps some AI predictions, but not every task

A University of Florida update on October 26 describes clinical AI comparisons using 22,254 hospital admissions. Models used structured records, radiology reports, chest X-rays, or combinations. UF reports gains of 21.4% for diagnosis prediction and 18.4% for medication prediction against the best single-source models. Structured records alone remained competitive for death-risk and length-of-stay predictions.

The work uses the first 24 hours of care from one hospital dataset. UF says diagnosis and medication scores remain at or below 0.6, below what real deployment requires. The reported percentages measure prediction improvements, not reductions in patient harm. The university posted its update on October 26, but that date does not make the underlying papers new publications.

University of Florida CTSI ↗
Read article →
05

Powerlattice tests a power chiplet on an AI accelerator stand-in

Powerlattice announced an integrated voltage-regulator chiplet demonstration on October 26. The company says the chiplet powered a silicon device it built to emulate the changing electrical loads of future AI accelerators. Its design combines the controller, power transistors, and inductors in one chiplet, placing voltage regulation closer to the processor.

This is a company demonstration on an emulated load, not an independently measured production accelerator or data centre. Powerlattice’s broader power-saving claims do not establish savings across a working facility. The release provides no independently verified customer deployment or fleet result.

Powerlattice via Business Wire ↗
Read article →

The watchlist

Signals that could change the read

An amended Anthropic prospectus with share count and price range.
Matched tasks that include corrections, elapsed time, and authorized actions.
The research groups’ conclusions and reproducible examination of the materials.
Tests at other hospitals and on production accelerators.
Across the desks Read the footnotes

Anthropic’s filing separates operating losses from financing charges. Model benchmarks and research announcements need their own checks.

Public marketsHotAnthropic filed its S-1. The company has not set the offering terms.
Computer workHotOpenAI's Terra comparisons still need tests on real work.
Scientific claimsHotOutside reviewers still need to examine OpenAI's claimed ETH refutation.
Clinical predictionWarmUF's retrospective comparisons do not show that patients benefit.

Editorial direction, not a financial index. Each signal is tied to this edition’s reporting.

Edition validated · 5 stories · 8 unique sources

About HotTea & our sources →