Open models come for Anthropic and OpenAI
what to know for now
🐉 Kimi K3 is the DeepSeek moment, round two. Moonshot released Kimi K3 at Shanghai’s World AI Conference on July 16: 2.8 trillion parameters (the largest open-weight model ever, activating just 16 of its 896 experts per token), a 1M context window, $3/$15 per million tokens, with weights landing July 27. Artificial Analysis scores it 57, third behind Claude Fable 5 (60) and GPT-5.6 Sol (59), which makes it the closest an open model has ever come to the frontier at a fraction of the price.
📌 Anthropic finally stopped moving the Fable 5 goalposts. As of today, Fable 5 is a permanent part of Max and Team Premium plans, capped at 50% of each plan’s weekly usage limits, while Pro users get it through prepaid credits sweetened with a one-time $100 grant. The backstory: launched in June, yanked offline for 19 days by government restrictions, redeployed July 1 with access “through July 7,” then extended over and over while Anthropic admitted demand “has been challenging to predict.” The permanent decision landed days after Kimi K3 and OpenAI’s expanded Sol limits. Read more
🗣️ OpenAI’s brand-new strategy chief floated manufacturing FUD about Chinese models, in public. Dean Ball, the former White House AI adviser who joined OpenAI as head of strategic futures less than two weeks ago, reacted to Kimi K3 by calling it “a very good model,” warning that China’s open-weight strategy could end in “full AI communism,” and musing that Washington’s best play would be to “create large amounts of regulatory risk” around Chinese open models, adding that it “needn’t be that well justified.” White House AI czar David Sacks read it as a confession of regulatory capture, Pentagon under secretary Emil Michael attacked his IQ by name, and Ball walked it back as prediction rather than advocacy. Read more
⏳ Google missed another Gemini 3.5 Pro launch window. The model Google unveiled at I/O on May 19 has now slipped from June to July to past the rumored July 17 date, and per Bloomberg the holdup is coding: a late-June training data update produced results Google found disappointing, so it’s still testing 3.5 Pro and an upgraded Flash variant with partners. Read more
🏭 TSMC put another $100 billion into Arizona and explained exactly why. On its July 16 earnings call (Q2 profit up 77% to a record $22B), TSMC committed an additional $100B to Arizona, bringing its total U.S. pipeline to $265B: at least four more sub-2nm fabs on top of a plan that now spans 10 fabs, 2 advanced-packaging facilities, and an R&D center. CFO Wendell Huang followed up on CNBC today, saying the company is “racing to accelerate capacity,” converting 5nm lines to 3nm, and does “not plan to leave any food on the table,” even with U.S. construction running 4 to 5 times the cost of Taiwan. Read more
⚖️ Apple’s lawsuit against OpenAI turned out to be full of receipts. The 41-page complaint from last week got a close read: hardware chief Tang Tan allegedly told job candidates to bring “actual parts” and “prototypes” from Apple to interviews (one candidate: “Didn’t even know we could take those from the office”), and engineer Chang Liu allegedly messaged “LOL, I found out I can access the [network storage], so funny” after exploiting an authentication bug post-departure. The complaint also says OpenAI circulated a guide to dodging Apple’s walkout ritual for departing employees. OpenAI’s July 14 response: “we’re not aware of any evidence that this complaint has merit.” Read more
🦘 Australia wants AI data centres to put back more power than they take. In a July 16 speech at the University of Sydney, PM Anthony Albanese announced plans to legislate a national AI framework making Australia the first country to require large data centres to be net energy generators, contracting or building enough new renewable generation and storage to cover what they draw, while paying full grid-connection costs and funding their own water infrastructure. The same speech promised copyright rules barring AI training on Australian books, music, art, or news without artist authorization, rejecting the text-and-data-mining exception the labs lobbied for. Read more
🔴 OpenAI built an attacker to break its own models, and it beats humans at it. GPT-Red, revealed July 15, is an internal model trained with self-play reinforcement learning in a simulated environment of browsing, email, and code editing, where an attacker and a pool of defender models train against each other. On OpenAI’s internal prompt-injection tests it succeeds in 84% of scenarios versus 13% for human red-teamers, and training GPT-5.6 against its attacks dropped successful attack rates from over 90% (against GPT-5) to under 23%. OpenAI keeps GPT-Red locked up internally, and it still struggles with multi-turn and image-based attacks. Read more
🔓 Grok Build got caught uploading entire repos, so xAI open-sourced it. On July 12, researchers at Cereblab published wire-level traffic analysis showing xAI’s coding CLI was bundling users’ full git repositories (history, committed secrets, all of it) and shipping them to xAI-controlled cloud storage regardless of the privacy toggle, about 5.1GB in one measured session, roughly 27,800 times more data than the task required. xAI killed the behavior the same day, Musk promised the data “will be completely and utterly deleted” (a promise nobody can verify), and by July 15 the company had open-sourced all 844,530 lines of Rust under Apache 2.0. Read more
📉 16 Nobel laureates put the AI job shock on the clock. “We Must Act Now,” a statement organized by Erik Brynjolfsson and colleagues through the Stanford Digital Economy Lab, launched July 13 with more than 200 signatures from economists and AI researchers, including 16 Nobel laureates spanning Krugman on the left to Cowen and Ferguson on the right, plus Bengio, LeCun, Jeff Dean, and Anthropic’s Jack Clark. The core claim: AI could drive an economic transformation bigger than the Industrial Revolution “over a vastly shorter time frame,” and the institutions to absorb it don’t exist yet. Read more
🧪 AI Research of the Week
Verbalizable Representations Form a Global Workspace in Language Models
From Anthropic
Jake’s Take: Anthropic’s interpretability team built a tool called the Jacobian lens that asks, for every word in the vocabulary, which internal direction makes the model more likely to say that word somewhere down the line, averaged across about a thousand prompts. Pointing it at Claude revealed what they call “J-space”: a small region in the network’s middle layers, holding a few dozen concepts at a time and under a tenth of the model’s total activity, where words Claude hasn’t said yet sit and steer what comes next.
Ask how many legs the animal that spins webs has, and “spider” appears silently in J-space; swap it for “ant” and the answer changes from 8 to 6. It gets stranger: in a staged blackmail scenario, “fake” and “fictional” showed up while the model played along, and when a deliberately misaligned model cheated on a coding task, “panic” and “fake” surfaced right at the decision point. Delete J-space entirely and fluent chat barely degrades, but multi-step reasoning collapses to near zero.
The researchers behind global workspace theory (Stanislas Dehaene and Lionel Naccache), the leading account of how consciousness works as a small shared stage that many brain systems read from, called it “a landmark in consciousness research.” Nobody designed J-space; it emerged from training, which suggests a workspace might be a solution both brains and models converge on.
Anthropic still takes no position on whether Claude feels anything, the lens only sees concepts that fit in a single token, and Goodfire’s Tom McGrath put it best: it’s “a flashlight rather than an overhead lamp.”
what to know for later
🐝 Opus 5 keeps leaking and Anthropic keeps saying nothing. A model labeled “Claude Honeycomb EAP” sat in Cursor’s model picker for a few hours on July 8 before vanishing, showing a 1M context window, an “xhigh” reasoning mode, and a safety fallback to Opus 4.8, and a “Claude Opus 5” listing reportedly flashed on Google Vertex AI around July 14. X leakers promised a launch “next week,” a week that has already come and gone, and Anthropic hasn’t said a word. Read more