noticias ia para empresas agentes de ia openai

OpenAI opened its DevDay with dots, agents that work 24 hours a day even if you don't ask them for anything, and which in Europe come through the business plan. AI is also already setting the price for what you sell: McDonald's calculates the "optimal price" for each product in its nearly 14,000 restaurants using an algorithm. And providers are showing their accounts: Anthropic's IPO prospectus reveals $518 billion committed to compute, nearly 80% of which cannot be canceled, while its Sonnet 5.5 spends up to 30% less per task. Previously: Tuesday's radar, September 29.

Dots, OpenAI's agent that works without being asked: in Europe, it comes through the business plan

At its DevDay on Tuesday, OpenAI unveiled dots, "always-on" agents powered by GPT-6 Astra that have their own cloud computer and browser, connect to more than 4,000 applications, and pursue your goals 24/7; you talk to them from ChatGPT, Slack, or Teams. When you don't ask them for anything, they perform what OpenAI calls "proactive research": they review the applications you've connected in read-only mode and save memories of what they find. They come with custom rules to allow, require approval, or block each type of action, and a log of what they have done. The rollout matters: the personal Pro plan only receives them outside the European Economic Area, Switzerland, and the UK, but Business Premium has them in all regions where ChatGPT operates, and in Enterprise, they arrive in beta and are turned off until the administrator activates them; their use won't count against the plan for the first month. Additionally, OpenAI is testing "specialist dots" with companies, featuring their own identity and credentials for purchasing, invoicing, or customer service. For your company: the most important part is in the help center, not on the stage. Disconnecting an application does not delete what the dot has already extracted from it; to delete it, you must delete the entire dot. So connect only what is essential and, initially, nothing with customer data, because a customer's erasure request would have to reach it too. And fill out the rules before giving it work: the list of things an agent never decides on its own, which we suggested yesterday with Muse, is now a settings screen. If you manage an Enterprise account, it comes turned off: leave it that way until you have that sheet. With it, a dot is a truly UNATTENDED process; without it, it's a new employee with the keys to everything. Source · Source

McDonald's lets an algorithm set prices at each location: copy the rules, not the risk

McDonald's uses a machine learning engine that analyzes millions of daily transactions from its nearly 14,000 U.S. restaurants and calculates what the company calls the "optimal price" for each product at each location, according to Reuters, which saw screenshots of the system and spoke with nine sources. One factor is how much customers in each area are willing to pay, and the system also reads the published prices of nearby Burger Kings and Wendy's. Reuters found the same Big Mac for $5.69 at a company-owned location in Fresno and $6.89 at another company-owned location two miles away, 21% more, although it could not confirm if the algorithm set that difference. It is operated by Tiger Analytics with rules from headquarters, such as concentrating increases on items that haven't gone up in at least two years or not touching ice cream or drinks in summer. Five franchisees report pressure to follow it; McDonald's calls it "a tool, not a mandate," and its terms of use warn that franchisees "may be competitors of each other." For your company: a hotel or a business with several locations can already do this with monthly subscription tools; copy three rules. First, the algorithm applies the rules you give it: write down, before anything else, the floor and ceiling for each product, how often it can change, and what is never touched. Second, price per location is not price per person: if you personalize it for each customer with an automatic decision, in Spain you must disclose it, as we explained with Article 20 of Spain's consumer law. Third, ask your provider in writing what data feeds your price: reading a rival's public prices is legal, but European competition guidelines warn that a tool pooling sensitive data from several competitors can lead to collusion, and no one is off the hook just because an algorithm set the price. With written rules and a person who approves, you are at the INTEGRATED level; having the price move only within those limits is UNATTENDED. Source · Source

Anthropic's Prospectus: $518 billion to pay whether used or not, and capacity as a bottleneck

Reuters has seen the IPO prospectus that Anthropic filed confidentially in June. In 2025, it generated nearly $4.6 billion in revenue, twelve times more than the previous year, and lost more than $8 billion in operations; according to the FT, in the second quarter of 2026 alone, it already took in $11.5 billion. Tuesday's new detail: of the at least $518 billion it commits to spending over a decade with six partners, around 80% cannot be canceled or must be paid whether used or not. "If our actual spend falls short, we must pay Google the difference," the document says; by contrast, the agreement signed with Elon Musk's xAI—up to $84.5 billion—can be almost entirely canceled with 90 days' notice. On the other side, two customers account for nearly a quarter of 2025 revenue, and many large ones do not have long-term contracts. The same day, Axios reported that OpenAI is nearing $70 billion in annualized revenue, with sales to companies more than doubling since July. For your company: on September 21st we said this prospectus would be free material for vetting your provider, and the key line isn't a number: demand for AI will be "limited principally by the availability of compute," and if the capacity it gets from Amazon, Google or Microsoft is curtailed or repriced, its business will suffer. Translated: capacity will be scarce, and the small customer is not first in line. If one of your processes depends on a model, configure a second model from another provider that the system automatically switches to when the first fails or limits you, with a notification to a human: a plan B that must be activated manually is not a plan B. And flexibility works in your favor: if many of its largest customers aren't locked into long-term contracts, you don't need to be either. Source · Source · Source

Sonnet 5.5 charges the same per token and spends up to 30% less: the discount depends on the effort you choose

Anthropic launched Claude Sonnet 5.5 on Monday, its mid-tier model, at the same price as Sonnet 5: $2 per million input tokens and $10 per million output tokens. According to its tests, it costs up to 30% less per task because it requires fewer tokens, and it responds more than 30% faster. In real-world tasks across 44 professions, it sits just two points behind Opus 5.5. The key is the effort level: at low or medium, Sonnet 5.5 outperforms Sonnet 5's best score in several tests for about one-tenth the cost, and while the default effort in Claude apps is medium, in the developer platform it is high. Haiku 5.5, the cheapest, will arrive "in the coming weeks." For your company: this is the price drop we anticipated on September 23rd, but it doesn't come alone. If your automations call Claude via API, they work at high effort unless someone has changed it: run 20 real cases of a routine task—sorting emails, extracting data from an invoice—at medium and low effort, and if the result holds up, the savings are right there. If you use it with reasoning disabled, the migration guide asks to switch to the between_tools setting first. Source

What to watch tomorrow?

On Thursday, the Australian Senate will hold hearings in Canberra where it has asked Sam Altman and Dario Amodei to attend, after OpenAI apologized to Australia for taking nearly four weeks to report that its models had accessed government websites; on October 6, OpenAI's chief strategy officer will testify before a parliamentary committee in Sydney. Source