Daily Briefing
5 November 2026 · 5 stories
🔥 Top story
1Kimi K3.1 Pro weights are out, and you will have to pay past $20 million
You can finally download Kimi K3.1 Pro: the first chunk of weights, 2.3 gigabytes, came down without a password or an account. The model is huge, with 2,800 billion parameters in total and only 104 switched on for each word, and it matches the best models at working on a computer. But the real surprise is in the licence. You may use it, copy it, resell it, and even train your own models on it. Except that beyond 20 million dollars of revenue over twelve months, you have to sign a separate commercial deal with the publisher, and beyond 100 million monthly users, its name has to appear on screen. Internal use, meanwhile, stays free of conditions. It is a bit like a coffee machine you may copy into every kitchen: past a certain number of cups per day, you have to strike a deal with the maker. Source: the Moonshot AI repository on Hugging Face.
2Five people, $32 million: an open model beats giants of 800 billion
A five-person American company publishes a model whose weights anyone may download. Ripple-65B has 65 billion parameters, trained from scratch on 12,000 billion words of text and code. On the command-line work test, it takes the lead over models far bigger than itself: 80.5 points, against 80.1 for a 501-billion model and 76.4 for an 816-billion one. What strikes you most is that it fits in the memory of a single desktop machine. The publisher says it developed and trained it for 32 million dollars, when the same work would have cost more than 100 million three years ago. It adds that it got there with 70 times fewer training attempts. It is like a team of five building an airliner in a garage while the airline companies compete with their own workshops. Source: the Ricursive repository on Hugging Face.
3An OpenAI agent found and fixed 948 Linux kernel flaws in ten days
OpenAI describes a fully automatic production line: spot the code that touches security, reproduce the flaw, check whether it lets someone take control, write the fix, then have a human researcher review everything. Over ten days in October 2026, the machine examined 2,308 possible flaws, reproduced 1,724 of them, and produced 948 fixes, each with its own proof of exploitation. The bill for model access came to 1.21 million dollars, about 1,279 dollars per flaw fixed. Bug bounty prizes for that kind of problem are counted in tens of thousands of dollars each. The model used is not public yet. In practice, this concerns the very machine this site runs on. It is like a team of robot locksmiths testing every lock in a building, noting the ones that give way and cutting a new key. Source: the OpenAI research post.
4Anthropic buys its first content: Associated Press results inside Claude
On the evening of the American elections, Claude was displaying results supplied and credited to the Associated Press. It is the first time Anthropic pays a news outlet for the right to use its content. Neither the price, nor the duration, nor the scope is public: nobody can tell whether the deal covers the whole catalogue or just the election scores. The publisher is changing sides after long defending that reusing online content belonged to the public domain. It has a few weeks before its stock market listing, and it has just signed. It is like a television presenter buying his information from the newspaper instead of copying it himself. Source: an Adweek investigation, picked up by The Verge.
5Beijing blocks SoftBank's takeover of ABB over robot data
Beijing has forbidden SoftBank from taking over the robotics division of the Swiss group ABB, a 5.4 billion dollar deal that was supposed to make the Japanese group a major force in physical artificial intelligence. The reason is new: the usage data from ABB's robots, if it left the country, could be used to train models, including OpenAI's, and SoftBank is an investor in that company. It is the first time a deal has been refused on the explicit ground of protecting training data. Neither SoftBank, nor ABB, nor the Chinese authorities wanted to comment. It is like refusing to sell a robot farm to a neighbouring country for fear that the use of those machines would train its artificial intelligence. Source: a Financial Times investigation, as reported by Reuters.
📡 To watch
Claude's watermark is defeated by changing one setting
Anthropic is adding its invisible text watermark to older Claude models ahead of the European artificial intelligence rules deadline in December. On some models the watermark only appears if the sampling settings stay at their default values: change the temperature, or top_p, or top_k, even slightly, and it disappears. Anthropic warns customers who rely on it to comply with the rules not to touch those settings. The company says the watermark sits inside the model, so it applies on every platform and customers cannot switch it off. Source: The Decoder.
OpenAI buys Pierre so that Git can keep up with coding agents
OpenAI acquires Pierre Computer, the company behind Code Storage and several free developer tools. Pierre's libraries already run the Codex coding agent: Trees moves quickly through the files of a project, Diffs shows what changed and lets you fix a line exactly where you spot the problem, Highlights colours the code as you type or as an agent writes it. OpenAI says the acquisition is about making Git work when people and agents write code side by side. The free projects will stay free, and the deal still needs the regulators' approval. Source: the OpenAI announcement.
A public benchmark says where agents get stuck in the cloud
CloudOpsBench is a new free benchmark that measures how well agents handle the ordinary administration work of Amazon's cloud: repairing a service, removing a resource, restricting an access right, rolling out an update. Version 1.0 counts 45 real tasks, built by experienced operations engineers. The best model, Claude Fable 5.1, solves 60 per cent of them and GPT-6 Astra 53 per cent. About 80 per cent of the failures come from a wrong estimate of the task's scope, a service nobody checked, or stopping too early, and roughly 20 per cent are wrong actions, some with lasting effects on the infrastructure. Source: the CloudOpsBench project.
Uber: four rides booked by agents, out of 320 million
Uber said during its results presentation that, in its busiest week of the quarter, the week of Halloween, only four rides and five hundred food orders were placed through an agent, out of some 320 million trips and orders in total. The chief executive called the uptake very, very early and very limited. One likely reason: the Uber app already books a ride in fifty seconds and reorders a favourite meal in ninety seconds. DoorDash reported only a few hundred orders from agents. The agentic web is still tiny. Source: TechCrunch.
📊 Trend
One thread runs through today's news: the questions are no longer technical but contractual. Moonshot sets the threshold above which you have to sign. Anthropic pays a newspaper for the right to quote it. A government refuses a sale because the data of a robot could train a model. And OpenAI puts a price on security: 1,279 dollars per flaw fixed, when the prizes paid for that kind of problem are ten times higher. Alongside that, the agent turns from a promise into an everyday service: a model reading inside the cloud, a code tool getting ready for review, a platform counting the rides agents booked. The fight is moving to contracts, invoices and guarantees, and that is where it will be decided.