noticias ia para empresas agentes de ia openai

OpenAI has launched GPT-6.1 Terra, a mid-tier model whose API costs half what its top model does and that comes close to high-end models on professional-work benchmarks. The Spanish Data Protection Agency (AEPD) has published ten rules for using AI assistants and agents: the agent drafts, a person approves, and no permission lasts forever. And the money is going to AI that sets prices: Guidewire will pay $1.7 billion, plus up to $400 million tied to targets, for UK-based hyperexponential, which prices insurance with AI. Picks up from the Tuesday, Oct 27 radar.

GPT-6.1 Terra costs half: the savings are only real if review time doesn't go up

OpenAI has launched GPT-6.1 Terra, which sits between the low-cost GPT-6 Luna and the pricier GPT-6.1 Sol and GPT-6 Astra and replaces GPT-5.6 Terra. On the API, it costs $1 per million input tokens and $5 per million output tokens—half the price of GPT-6.1 Sol and Anthropic's Claude Sonnet 5.5 ($2 and $10). OpenAI is marketing it for high-volume office work: drafting proposals, interpreting blueprints, filling out forms and running desktop applications. In OSWorld 2.1, a benchmark where AI uses computer programs for multi-step tasks, it achieves 74.2% at $1.19 per task, compared to 57.1% and $4.76 for its predecessor; Claude Sonnet 5.5 remains ahead with 83.9%, but at $5.74 per task. It is already available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu plans. Sanchit Vir Gogia, chief analyst at Greyhound Research, offers the counterpoint: in an illustrative coding workload, 5.4 extra seconds of human review per task erase the inference savings. For your business: if you run high-volume, low-complexity processes such as invoices, forms or document sorting, this is the model to test. But don't just compare prices per million tokens: take 20 real tasks, run them through your current model and Terra, and time how long a person takes to review each result. If review time increases, the savings are only on paper. And if the error rate drops as much as OpenAI promises—4.3% in its internal tests at maximum reasoning—that process could move from checking every result to spot checks: that's the jump from ASSISTED to UNATTENDED. Source

Spain's data regulator sets the rules for AI agents: the agent drafts, a person approves

The Spanish Data Protection Agency has published "AI Assistants and Agents: A Ten-Point Guide for Users." Its starting point: "When an assistant only responds, an error is a wrong answer. When it acts, it is a sent email, a shared file, a purchase, a contract, or a consent given in your name." It proposes five questions before delegating: what can it see, what can it remember, what can it do on my behalf, where does it send data, and does it affect anyone else? It recommends granting only the essential permissions for each task, without permanent authorizations, citing a case where a user clicked "always allow" to sell an item: the agent closed the deal and gave the user's address to the buyer, who then showed up at their house. It warns that a website, an email, or a calendar invite can hide instructions for the agent, and that telling it what not to do is no substitute for permissions or confirmations. At work, it advises using only organization-authorized assistants and not connecting professional email or documents to a personal AI account. And it says providers should offer permissions that tell looking apart from acting, a history of what the agent did, and an immediate way to stop it. For your business: the guide is written for users, but it gives you a ready-made internal policy. Publish a list of authorized assistants, explicitly prohibit connecting company email to personal accounts, and use the controls the AEPD asks of providers as a shopping list: if a tool doesn't separate looking from acting or doesn't keep a history, don't let it work unsupervised. Source

Guidewire pays $1.7 billion for hyperexponential: the money goes to AI that sets prices with your data

Guidewire, which makes policy, claims, and billing software for many P&C insurers, has signed a deal to buy hyperexponential, a London-based company whose platform is used by insurers, reinsurers, and underwriting agencies to build pricing models and evaluate complex risks. It will pay $1.7 billion in cash and up to $400 million more if recurring revenue targets are met. Hyperexponential was already a Guidewire partner, brings about 300 people, and its tools are involved in decisions regarding more than $75 billion in premiums annually. The deal is financed by a $1 billion loan from Bank of America and Citibank plus existing cash; Guidewire expects to close in the second quarter of its fiscal 2027. The pitch is what they call "agentic underwriting": AI agents that evaluate risk, price it, and decide alongside underwriters and actuaries. For your business: the big money goes to AI tied to the decisions that make money, like pricing or accepting a customer, and plugged into the systems where your data already lives, not to generic chat. If quoting or pricing in your business depends on one or two people and a spreadsheet, that's your use case: let the AI prepare the price using your data while the expert retains the sign-off. That's the INTEGRATED rung, the one that pays. Source

What to watch next

Microsoft reports quarterly earnings today, Wednesday, Oct. 28: we'll see what its voluntary buyout program costs and how fast its enterprise AI business is growing. And it's still unclear whether Anthropic will answer with price cuts of its own against a model that costs half as much as Sonnet.