noticias radar ia agentes de ia costes de ia automatizacion

Yesterday's radar was about the cost of each model call; today's is about who owns the tool you already have integrated into your workflow. In twenty-four hours, SpaceX has closed the acquisition of code editor Cursor for $60 billion, and Anthropic has reported the first quarter with operating profit for an AI lab, on track for an IPO in October. In parallel, an uncomfortable fact for those about to switch models: the biggest cost reduction of the day didn't come from a new model, but from the software surrounding it. And in Spain, the Community of Madrid is putting money on the table for SMEs to take the leap.

Your code editor changed owners yesterday: Cursor now belongs to SpaceX

The merger by which SpaceX acquires Anysphere, the company behind Cursor, became effective on August 14 according to the filing submitted to the regulator: an all-stock deal with an implied value of $60 billion, in which Cursor investors receive 389.3 million SpaceX Class A shares. Signed on June 16, it closed in less than two months; Cursor becomes a wholly-owned subsidiary and its team joins SpaceXAI alongside Grok Build, Grok Bot, and the Grok API. For your company: if your team programs with Cursor, your provider is different today, and the terms are set by the buyer, not the person who sold you the subscription. Check whose name the contract is in, what it says about using your code to train models, and where what you upload is stored; if any of that changes, you'll find out through a terms update email, not a call. And the general lesson applies to any tool you put into a process: a buyout closes in two months, so be clear on how you would export your configuration and how long it would take to replace it before you need to. Source

The first AI lab to make money goes public in October, and that hits your invoice

Anthropic closed the second quarter with $10.9 billion in revenue, more than double the previous quarter, and with the first operating profit in its history—about $559 million according to investor materials—something no leading lab had achieved. Its annualized rate jumped from about $9 billion at the end of 2025 to $47 billion by mid-May, supported by more than a thousand enterprise accounts spending over a million dollars a year on its API alone; it filed its IPO draft confidentially in June and aims to list as early as October with a target valuation of two trillion dollars. The figure should be read carefully: operating profit does not discount debt interest or taxes, and the company itself warns that profitability may not be sustained when full data center spending kicks in. For your company: a publicly traded provider stops setting prices to gain market share and starts setting them to defend margins for the market. The practical translation is simple: don't sign an annual commitment at today's price without an exit clause, and have a measurement of what it would cost to move your automations to another model—not to do it, but to know if they can squeeze you. There is no stock advice here: there is a date in October and one of your providers on the other side. Source

Savings aren't in switching models, they're in the software around it

Writer published Palmyra X6 along with a rewrite of its "harness," the layer that orchestrates the model and decides what context to pass it at each step. The interesting data isn't the proprietary model—52% less cost, 48% faster, and 10% better quality—but what happened when running the new harness with external models: with Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, and GLM 5.1, the cost per task dropped by 41% on average, from 14,200 to 8,800 tokens and from 21 to 12 cents per task, without touching the model. For your company: half of your AI bill isn't decided by the model you chose, it's decided by how many times you talk to it and how much text you resend in each call. Before jumping to the trendy model, open your automation and count: how many calls it makes per task, how much context it carries in each, and how many times it resends the same information it already sent before. Cutting that repeated context and grouping steps usually saves more than any provider change, is reversible, and doesn't tie you to anyone. Source

Madrid puts 6 million to help regional SMEs use AI, and its catalog is useful even if you aren't in Madrid

The Community of Madrid has authorized 6 million euros through 2028 to renew its agreement with IndesIA and maintain the AI4Pymes Madrid platform, which added to the 7.6 million in Next Generation funds from the first phase brings the total investment to 13.6 million. The platform is aimed at micro, small, and medium-sized enterprises with fiscal residency or activity in Madrid and offers five specific tools: operations evaluation, virtual assistant creation, public tender processing, search for public grants, and digital file conversion. For your company: keep that catalog of five even if you aren't in Madrid, because it's the list of processes already considered solved with AI in a Spanish SME—reading and classifying documents, setting up an assistant, preparing a tender, searching for calls. If you are in the region, try them before paying for an equivalent tool; if not, use that list as a guide to decide where to start and check what similar line your own region has, because almost all of them have one. Starting with a process others have already automated a thousand times is cheaper than inventing your own. Source

What to watch tomorrow?

Today, Sunday at 16:00 UTC, DeepSeek's new rates take effect, so you'll see the first bill with the new price tomorrow, Monday. And on the 23rd, data generated since December 29 will be deleted in Manus: if someone on your team uses it, it's time for a backup this week.

Watch the 1-minute video