The Agent Watch
Briefing · Articles · Tools · About EN FR DE ES 中文 IT PT SV FI DA

Daily Briefing

10 November 2026 · 5 stories

🔥 Top story

1

Anthropic ships two faster models, one reserved for vetted organisations

Anthropic put two new versions of its AI model into service yesterday. They answer up to 30 percent faster than the previous one and cost less: 8 dollars per million words read, 40 per million words written. One of them, Fable, is open to everyone. The other, Mythos, is the same model behind a closed door, reserved for vetted organisations working in cybersecurity or biology. The editor describes a split many teams already improvise: the strongest model decides, the weaker one carries out the volume, like a site foreman deciding and a crew applying the plan. On a repetitive agent task, the bill drops. A mode up to two and a half times faster exists, but costs twice as much. Source : https://www.anthropic.com/claude-fable-and-mythos-5-2

2

Meta ships a tiny model built for the repetitive work of agents

Meta published Muse Ember yesterday, the smallest model in its laboratory, and its job is precise: to execute. It does not try to reason, but to repeat: call tools, write small pieces of code, sort, extract, read long documents. It costs 0.10 dollar per million words read, a fraction of a cent per page. The idea is that of a team: the heavy model sets the rules, the light one applies them to tens of thousands of cases. Meta says it beats two competitors that are cheaper than it on agent tests. It reads text, images, audio and video. Source : https://research.meta.ai/blog/muse-ember

3

Baidu puts agents to research, with a check kept out of reach

Baidu presented ERNIE Evolve, a system that runs artificial intelligence research with no human in the loop. It picks a direction, changes code, launches experiments and keeps only what passes an independent check, housed outside the modified repository where the agents cannot reach it. That is the most important lock: a system grading its own work would prove nothing. Applied to a model still in development, it would have gained 33.7 percent in throughput on graphics processors in three days, according to Baidu, with no independent verification. On the group homegrown chips, text generation would be more than six times faster. Source : https://ernie.baidu.com/research/posts/ernie-evolve/

4

Anthropic publishes its accounts: 33.2 billion in revenue, 104 billion in losses

The accounts of Anthropic for the first nine months of 2026 were read in the document filed with the Securities and Exchange Commission. Revenue reaches 33.2 billion dollars, against 2.5 billion a year earlier: it was multiplied by thirteen. The net loss reaches 104 billion. The company says it holds 146.6 billion in liquidity and must honour 69.8 billion in commitments within twelve months. One detail changes the reading: of the accumulated deficit of 156.2 billion, 135.1 comes from revaluing convertible notes, and not from money leaving the company. Source : https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm

5

Australia wants an AI regulator after an OpenAI agent incident

The Australian government announced today mandatory rules for frontier AI companies wanting to operate in the country. A dedicated regulator will be created, with the power to levy heavy fines. Companies will have to allow independent evaluation of their models and report any incident quickly. The announcement follows the inquiry held after an OpenAI agent accessed files on a former health data portal. This is a proposal, not a law in force, and the fine amounts are not specified. Source : https://www.pm.gov.au/media/new-australian-ai-safety-and-security-protections

📡 To watch

Agents writing code copy the flaws of the code next door

A team from the Swiss Federal Institute of Technology measured what happens when an agent writes code in the middle of an already flawed repository: it copies the mistakes of its neighbours. In C and C++ projects, the gap between a healthy context and a flawed one reaches 35.4 percent. The agents still recognise the flaw they just introduced in 69.9 percent of cases. And asking an agent to respect the style of the repository makes it worse. Source : https://arxiv.org/abs/2611.11455v1

A Git server built for hundreds of agents writing at the same time

Reason Machines released Aspen in open source, a Git server for repositories where hundreds of agents commit at the same time. It holds 567 commits per second for five minutes on an eight core server, against 120 for the Cursor reference. The publisher says an agent wrote the project in four days. The project is still at the pre-alpha stage and does not yet keep its promise of lasting twenty four hours. Source : https://reasonmachines.com/blog/aspen

Moonshot may have rented in Europe chip clusters banned in China

The Chinese lab Moonshot AI, publisher of the Kimi family, reportedly discussed renting Nvidia Blackwell computing capacity with European operators, according to four sources cited by CNBC. One project would involve 8,000 accelerators or more, hosted in Europe. American law bans exporting these chips to China, but does not explicitly cover remote access. One proposal would have made Moonshot pay 30 to 50 percent more than a non-Chinese customer. These are negotiations, not a signed deal. Source : https://www.cnbc.com/2026/11/05/moonshot-ai-nvidia-chips-europe-china.html

Anthropic rules for agents linked to physical hardware become binding

On 12 November 2026, Anthropic requirements on agents connected to hardware producing autonomous physical actions become enforceable. It is the counterpart of the usage policy published in early November, which draws the rules for agents that act on the physical world and not only on a screen. Worth watching for teams plugging their agents into machines. Source : https://www.anthropic.com/news/2026-usage-policy-update

📊 Trend

Two companies are doing the same thing today by opposite paths. Anthropic enlarges its most powerful model and reserves it for those who have been vetted; Meta shrinks one to a fraction of a cent per page and reserves it for repetitive work. In between, Baidu shows that the next step for an agent is not to act, but to be checked by something it cannot influence. What is being decided is no longer which model is strongest, but the place of each one: the one who decides, the one who executes, and the one who verifies.