GPT-5.6 rumors stir as Anthropic pushes a more “honest” Opus 4.8
🤔 “Why are there so many new models?”
Share Handy AI with your coworkers and friends to help them understand the crazy world of modern artificial intelligence and make the right decisions.
what to know for now
🤖 GPT-5.6 is leaking out of OpenAI’s own logs ahead of a rumored June launch. A gpt-5.6 identifier briefly surfaced in Codex traces, internal codenames (ember-alpha, beacon-alpha, iris-alpha) are floating around, and the chatter points to a 1.5M-token context window aimed at full codebases and book-length documents. Nothing is official, no model card, no benchmarks, but prediction markets put a public release before June 30 at 80 to 89%. Read more
🧠 Anthropic shipped Claude Opus 4.8. The May 28 release pushes agentic coding from 64.3% to 69.2%, multidisciplinary reasoning from 54.7% to 57.9%, and knowledge work from 1753 to 1890, all at the same price as 4.7. The headline features are effort control on claude.ai, dynamic workflows in Claude Code that spin up parallel subagents for large jobs, and a fast mode that runs 2.5x faster and three times cheaper than before. Read more
💰 Anthropic raised $65B at a $965B valuation and is now the most valuable AI startup on earth. The Series H, led by Altimeter, Dragoneer, Greenoaks, and Sequoia, nearly triples February’s $380B mark and edges past OpenAI’s $852B. The justification is a $47B revenue run rate, up from $30B earlier this year, powered mostly by Claude Code. Read more
🏛️ OpenAI finished its recapitalization, and the nonprofit it spent a year trying to escape now controls the whole thing. The OpenAI Foundation sits on top of the public-benefit company, holds roughly 26% of the equity (about $130B today), and Altman wants it to become the largest nonprofit in history. It opens with a $25B commitment to health and curing disease plus AI resilience, and plans to spend at least $1B over the next year across life sciences, jobs, and community programs. Read more
🧬 Chan Zuckerberg Biohub released a protein “world model” that designs new drug molecules in hours instead of years. Built on the fourth generation of evolutionary scale modeling (ESM), the system predicts protein structure, maps how proteins behave, and designs fresh binders that lock onto specific targets. In lab tests, binders it designed for cancer and immune targets actually reactivated immune cells. Read more
📉 Uber’s COO said the quiet part out loud: the AI bill is real, the payoff is not yet. Andrew Macdonald admitted it’s “very hard to draw a line” between rising token spend and shipped customer value, after the company burned its entire 2026 Claude Code and Cursor budget in four months. The numbers are wild in both directions: 95% of Uber engineers touch AI tools monthly and 70% of committed code is AI-generated, yet none of it maps cleanly to product wins. Read more
🧪 AI Research of the Week
Leveraging Large Language Models to Improve Precision in Randomized Controlled Trials
From University of Michigan / WPI
Jake’s Take: Randomized controlled trials are the gold standard for proving a drug or intervention works, and they are typically crazy expensive because you need a lot of participants to see a signal through all the noise. This paper feeds LLM predictions as an additional layer on top of this analysis (not as a replacement for the real trial data), which seems to tighten the estimate of the treatment effect. It works best exactly where trials have struggled, in studies that lack good predictive variables or that lean on messy text data (the model can read better than a regression can).
The so-what is sample size. If an LLM-assisted analysis squeezes the same statistical confidence out of fewer patients, you can run trials that were previously too small, too rare, or too costly to attempt, which is the whole bottleneck for rare-disease and under-funded research. It pairs neatly with other stories from this week: Biohub designing molecules and OpenAI’s foundation pledging $25B at disease both assume the testing pipeline can keep up, and this is one way it does.
what to know for later
🔁 The ex-DeepMind crowd is racing to build AI that improves itself without human data. David Silver, the AlphaZero architect, left DeepMind and raised $1.1B at a $5.1B valuation for Ineffable Intelligence, a “superlearner” meant to discover knowledge through reinforcement learning rather than human examples, backed by Sequoia, Lightspeed, Nvidia, and Google. A parallel outfit, Recursive Superintelligence, pulled $500M to build a model that finds its own weaknesses and rewrites itself. Read more