On Thursday, October 9, 2026, British bank HSBC and Ant Digital, the fintech arm of Chinese giant Alibaba, ran a historic test together: an AI agent discovered an online service, used it, started a payment of less than two dollars and closed the transaction end to end, with no human in the loop. To do that, they combined HSBC's tokenised deposit platform with Ant Digital's private Jovay blockchain. No commercial launch is planned yet, but it is the first time an intelligent agent has both initiated and settled a real payment while honouring bank controls in real time. For most people, it is as if your home assistant could now, on its own initiative, pay the pizza delivery, the Netflix subscription or the electricity bill, while letting the bank check every move. Source: https://www.prnewswire.com/2026/10/09/ant-digital-hsbc-agent-payment/
On Tuesday, October 6, 2026, French startup Mistral unveiled Le Chonk, its new AI model nicknamed internally Le Chonk: one trillion parameters in total, fifty-two billion used at each question, trained on nearly four thousand Grace Blackwell chips in a European data centre. Its specialty is cybersecurity. On the Cybench benchmark, which measures the ability to solve computer-hacking challenges, it scores ninety-three per cent, the world's best. The model's open weights will be published at the end of October on Hugging Face. For most people, it is as if a European company had finally launched the world's best digital detective, able to find the hidden flaws in software before attackers use them, with an open manual that anyone can audit. Source: https://mistral.ai/news/le-chonk-cybench/
On Tuesday, October 13, 2026, Causa Prima, a German-American startup, closed a twenty-million-dollar funding round led by Base10 Partners. Its specialty: making companies' AI agents talk to each other directly, so they settle disputed invoices, negotiate payment terms or schedule advances on their own, without a human re-reading every exchange. The round also includes QED, FJ Labs, Creandum and Kfund, for a total of thirty million raised since founding. For most people, it is as if two robot accountants, in two different companies, could now agree between themselves on a disputed invoice, sign a deal and trigger the bank transfer, without anyone picking up the phone. Source: https://techcrunch.com/2026/10/13/causa-prima-base10/
On Tuesday, October 13, 2026, Loops, a startup co-founded by Ari Nazir and İlker Selim Zorluoğlu, raised three and a half million dollars from a16z speedrun and NTT Docomo. Its mission is an operating system for online shopping: a single AI guides the shopper from the first search to the payment, through product choice and virtual try-on. Seventy clients or pilots already use it, with a conversion rate of sixty-four per cent from pilot to paying. The Turkish group A101 reports an eighty-three per cent lift in online purchases thanks to Loops, a figure to be taken with caution because it has not been measured scientifically. For most people, it is as if a personal assistant accompanied every visitor on your shop site, from the home page to the final click, asking the right questions and removing hesitation. Source: https://speedrun.a16z.com/portfolio/loops/
On Tuesday, October 13, 2026, Google began the global rollout of Gemini 4 Argon, its new AI model, available in the Gemini API, AI Studio, paid subscriptions and even the Google search engine. The entry price is two dollars per million input tokens, ten for output, with a ninety-five per cent discount on cached responses. On coding and automation benchmarks, it scores the best on the market, and it is also the first model to handle one million tokens of output context. The rollout strategy mirrors cybersecurity software: first to defenders, then open to all. For most people, it is as if Google had launched a new version of its search engine, but augmented, able to answer much longer questions and complete complex chores in seconds. Source: https://blog.google/technology/ai/gemini-4-argon-rollout/
Meta joins the frontier race with a model that remembers
On Sunday, October 12, 2026, Meta unveiled Muse Wonder, its first so-called frontier model. It scores fifty-six on the Artificial Analysis AI Index, an independent test that places the model at the level of GPT-6 Astra, behind Claude Opus 5.5 and neck on neck with GPT-6 Sol. Its specialty is a persistent memory that can keep days, weeks or months of history. It is the first time an independent test confirms that Meta is back in the race after fourteen months of stagnation. Source: https://techcrunch.com/2026/10/12/meta-muse-wonder/
The West's DeepSeek will release its weights in late October
For the record: American startup Reflection will publish in late October 2026 the open weights of its Beam model, five hundred and one billion parameters of which twenty-three billion are active per question, under the Apache 2.0 licence. Beam was trained on twenty-three trillion tokens with a massively reduced compute budget, about one third of what China's GLM 5.2 cost for similar scores. Ollama already ships it. First open sub-agent on a par with Chinese models, worth watching closely. Source: https://reflection.ai/beam-open-weights/
Nvidia loses half of China, Huawei rises fast
On Tuesday, October 13, 2026, according to IDC, Nvidia's market share for AI data-centre chips in China fell to thirty-four per cent in 2026 while Huawei climbed to twenty-eight per cent. Nvidia's boss, Jensen Huang, puts at about fifty billion dollars the Chinese market handed over to Huawei because of US export controls. For most people, it is as if a company that had dominated an Asian market in a few years was caught up and overtaken by a local rival boosted by regulation. Source: https://www.reuters.com/2026/10/13/idc-nvidia-huawei-china/
Mistral will release its giant model's weights on October 27
For the record: French startup Mistral will publish on October 27, 2026 the open weights of its Le Chonk model, one thousand and fifty billion parameters in total, fifty-two billion active per pass, trained on three thousand eight hundred Grace Blackwell chips. Already available as a preview on the Mistral API since October 6. First European AI model of this size to be open-sourced, a sectoral event to watch closely. Source: https://huggingface.co/mistralai/le-chonk/
📊 Trend
Three big lessons from the day of October 14, 2026. First, money changes hands: for the first time, an AI agent both initiates and settles a real bank payment with no human in the loop, while honouring the bank's real-time controls. Second, security changes continent: European model Mistral Le Chonk scores ninety-three per cent on the world Cybench cybersecurity test, ahead of OpenAI and Google. Third, commerce changes engine: two startups, Causa Prima and Loops, raise funds so that AI agents negotiate and sell on behalf of humans. The day's big convergence: everywhere, intelligent agents act on behalf of humans, whether to pay, to defend or to sell. And the big question now: who watches these agents, and at what moment can we stop them? That is exactly what the Agent Wealthy multi-surface agent security audit vertical has been waiting for.