noticias ia para empresas automatizacion de procesos agentes de ia

Today's news is all about the invoice: Anthropic lowers the price of its most expensive model on the same day an independent analysis proves that price per token no longer predicts what it costs to finish a job. Behind the scenes, the infrastructure provider goes public with 85% of its revenue tied to two clients, Toyota puts a figure on robotizing its factories, and twenty-two governments ask the UN for IA not to decide without a human in the loop. This follows the radar from Tuesday the 22nd, where the Government banned spending its 600 million grant on licenses.

Cheap AI is expensive: price per token no longer tells you what a job costs

Anthropic launched Opus 5.5 yesterday and lowered its starting rate from $25 to $20 per million tokens, with similar cuts across other concepts and a model that, according to the company itself, outperforms Fable in many tests. The same day, it became clear why that figure is no longer enough for budgeting: an independent analysis by Artificial Analysis on Grok 4.7—which maintains prices of $2 and $6 per million, among the cheapest on the market—found that it consumes 81,000 output tokens per task compared to 27,000 for GPT-6 Astra Max, three times as many. The result: each completed task costs about $3.74 with Grok 4.7 xHigh and $1.99 with GPT-5.6 Sol Max. The model with the cheapest rate ends up being the most expensive to use. For your company, this changes the unit you negotiate with and decide by: stop comparing providers by price per million tokens and set up a measurement this week of cost per successfully completed task across three or four of your real workflows—a customer response, a case summary, a recurring report—tracking consumed tokens, retries, and how many times a person had to intervene. Without that number, you don't know if switching models saves you money or costs you, and with it, the renewal conversation stops being a rate dispute. Source

Nscale goes public: 103 billion in contracts and two clients accounting for 85%

Nscale filed to list on the New York Stock Exchange seeking $3 billion at a valuation of approximately $35 billion. The prospectus provides what was missing when we reported in the September 5th radar that it was negotiating 3.5 billion in convertibles before going public: revenue of 140.6 million in the first half of 2026—compared to 10.4 million a year earlier—and net losses of 1.02 billion in that same semester. Of the 103 billion in accumulated contracts, 43.8 billion are from Microsoft and 44.6 billion from Anthropic: two clients, about 85% of the contracted revenue. And the new detail is the most uncomfortable: the agreement with Anthropic is conditional on Nscale securing financing, and the client reserves the right to terminate if it fails to meet milestones that the document itself describes as rigorous. For your company, this isn't a stock market story; it's the map of your supply chain: your AI provider's provider lives off two contracts that could be canceled. Do the reverse exercise this week—how many of your automated processes depend on a single model, and a single site where that model runs—and write down next to it what you would do next week if that provider raised the price by 40% or stopped service. If the answer is "redo the work," you are not yet at the INTEGRATED level. Source

Toyota puts a price on the automated factory: 400,000 robots and 6.4 billion a year

Toyota explained to its investors that extending automation to its plants, group companies, and main suppliers could require about 400,000 robots and an annual expenditure close to 1 trillion yen—about $6.4 billion—starting in 2028. The figure includes both replacement of existing machines and new installations, mixing humanoid robots with conventional industrial systems, automated logistics, and shared workstations between humans and machines. For your company, the useful part isn't the amount but the timeline and the order: the world's largest automaker places the heavy spending in 2028 and dedicates the two prior years to deciding what to automate and what to connect. If you have a workshop, warehouse, or assembly line, that is exactly your deadline to perform the boring inventory no one does: which of your machines expose data—serial port, OPC-UA, Modbus, an API—and which are black boxes that will need to be replaced so any agent can read them. Start by reading, not writing: counters, stops, cycle times. Those who reach 2028 without knowing what their plant measures will buy robots they won't be able to coordinate. Source

Twenty-two governments call at the UN for AI not to decide without a human present

During the United Nations High-Level Week, leaders from 22 countries and institutions—including Pedro Sánchez and Ursula von der Leyen, along with Germany, the Netherlands, Canada, Singapore, and the United Arab Emirates—signed a declaration calling for common standards and maintaining that "AI must remain under human direction, supervision, and control." The text proposes studying an international institution to set standards and facilitate verifications, transparent safety protocols with mandatory testing before deployment, and independent model evaluations. It is non-binding and no company has signed, so nothing changes tomorrow; what it does anticipate is the direction of the paperwork to come. The technical counterpoint was published this month by a researcher on arXiv under the name loopjacking: he demonstrated that in several agent frameworks, a reviewer can be tricked into approving one operation while a different one is executed, reproducing this in versions of Agno AgentOS and LangGraph. For your company, the joint reading is that "human supervision" isn't just putting someone to click the accept button: if what is approved and what is executed are not literally the same, the signature is worthless. Check your automations to ensure the approval screen shows the exact action to be launched and that nothing can modify it between the click and execution, and keep a log of what was approved. Source

What to watch tomorrow?

Nscale's opening price and whether investors accept 85% of revenue from two clients: it marks the cost of compute that you end up paying for. And the Sonnet 5.5 and Haiku 5.5 versions that Anthropic promises "in the coming weeks," because those are the ones that truly move the invoice for an SME.