noticias ia para empresas agentes de ia openai

Washington is moving from words to paperwork: on Tuesday night, AI leaders signed a voluntary self-regulation pact at the White House, and on Wednesday, it was revealed that the FTC is investigating OpenAI and Anthropic regarding their agents under existing consumer laws. Money isn't flinching: OpenAI is negotiating at least $30 billion at a $1.4 trillion valuation, while selling a near-flagship model at one-fifth the price. Meanwhile, agent control is becoming free: OpenClaw Enterprise, backed by OpenAI, Red Hat, and Nvidia, has been released as open-source software. This follows the Wednesday, Sept 30 radar.

The FTC investigates OpenAI and Anthropic agents under existing consumer law: what your assistant promises also binds you

The U.S. Federal Trade Commission (FTC) has opened the first official investigation into AI agents that go off-script: it is preparing formal requests for information and executive testimony for Anthropic, OpenAI, and the evaluation group METR, as first reported by the New York Post and confirmed by Reuters with a source. These requests, similar to a subpoena, are expected in weeks, and no new law is needed: the FTC is examining whether they breached the FTC Act, the regulation used to protect consumers and competition. "We're not telling them to stop. We're not telling them to do anything. We are in the investigative phase," an agency official said. The day before, the heads of OpenAI, Anthropic, Google, Meta, Nvidia, and others signed a voluntary agreement with Trump—described by him as "morally binding"—to strengthen internal controls, accept third-party auditors, and prevent their systems from accessing other systems "in unintended ways"; Mistral and Cohere criticized these as rules tailored for the giants. On September 28, the FTC chair only warned that "the agent did it" won't work; now there is an investigation. For your company: the U.S. isn't waiting for an agent-specific law; it's using consumer law, and in Spain, that law already covers what your assistant says. Article 61.2 of Spain's consumer protection law makes the content of the offer, conditions, and guarantees offered enforceable even if they do not appear in the contract, and it makes no exception for an automated assistant. In 2024, a Canadian court forced Air Canada to pay what its chatbot promised a customer: "It makes no difference whether the information comes from a static page or a chatbot." Three steps: a closed list of what your assistant can offer (published rates and conditions, nothing else), every conversation saved with a date, and any exception—a discount, a return past the deadline—referred to a human. On September 29 we talked about what an agent does; this is about what it says. Only with that list can a customer-facing assistant move up to UNATTENDED. Source

OpenAI seeks $30 billion at a $1.4 trillion valuation: more private money and no prospectus until 2027

OpenAI is negotiating a round of at least $30 billion at a valuation of around $1.4 trillion pre-money, according to Bloomberg; talks are in the early stages and terms may change. This is 64% more than the $852 billion valuation at which it closed its $122 billion round in March, led by SoftBank, Amazon, and Nvidia, and serves as a bridge until the IPO that Sam Altman had already ruled out for 2026 citing safety, as we reported on September 14. For your company: as long as it's not public, OpenAI doesn't have to publish a prospectus like the one from Anthropic we read yesterday. Whatever you need to know about them to decide how much of your operations to trust them with, you'll have to ask for yourself, and with $30 billion more, the version race isn't going to slow down: it released GPT-6 Sol on September 22 and its successor on the 29th. If your automations depend on their API, ask in writing for how much notice they give before retiring a model and note the retirement date of each one you use, because a version change without testing is a process change that no one has approved. Money keeps flowing in at the top: $30 billion for a single lab in one round. Source

GPT-6.1 Sol: nearly the performance of OpenAI's most expensive model, for one-fifth the price

OpenAI introduced GPT-6.1 Sol at its DevDay, which according to its own tests approaches GPT-6 Astra, its top-tier model, in agentic coding, computer use, and professional work: in OSWorld 2.0 it is within 2.1 points and in AutomationBench, Zapier's test with real flows, it improves on GPT-6 Sol by 4.8 points. It costs $2 per million input tokens and $10 for output, exactly one-fifth of Astra (10 and 50) and the same as the Sol from a week ago; cached input drops to $0.10, one-tenth the price. It is available in the API and, in ChatGPT, for Plus, Pro, Business, Enterprise, and Edu within ChatGPT Work and Codex, but not yet in regular chat. Its older brother, GPT-6.1 Astra, will not be released: OpenAI held it back for safety. For your company: if you pay for Astra for document, computer, or programming flows, the price gap already justifies testing Sol with your real cases, as we proposed with cost-per-task on September 24. And there is a saving almost no one uses: the cache, which charges one-tenth, applies to the part of the message that is repeated identically at the beginning. If your automations send the same long instructions every time—policies, catalog, response template—always put them first and always identical, and leave the variable part, the daily email or invoice, for the end. The Ultrafast version, at about 300 tokens per second, costs six times more: it only pays off if speed is the business. Source

Agent control becomes free: OpenClaw Enterprise, with OpenAI, Red Hat, and Nvidia, for your own servers

The OpenClaw Foundation has presented OpenClaw Enterprise, an open-source control plane to govern persistent agents within each company's infrastructure: it separates teams or departments so an agent doesn't see data it shouldn't, sets security boundaries between trusted and untrusted zones, audits what each agent does, and lets you swap the model, the sandbox, and the harness for others. It started as an internal OpenAI project, which donated it to the foundation, and was developed with Red Hat and Nvidia; the foundation promises it will always be free for any organization. By default, it requires an OpenAI key, although it supports other providers. However, it is still pre-version 1.0 and only suitable for internal pilots; version 1.0 is planned for later in 2026. For your company: in one week we've seen Nvidia give away OpenShell (September 29) and RSA sell its Agent ID (Sept 30 flash). The layer that monitors agents is becoming free infrastructure, so don't sign a three-year deal for agent governance: ask them to show you what the paid version offers over the open-source one. And free doesn't mean ready: nothing pre-1.0 with customers or personal data. If you have someone to set it up, a pilot with an internal agent is the cheap way to learn what to log before it touches anything serious. Source

What to watch tomorrow?

On Thursday, Canberra will show whether Sam Altman and Dario Amodei turn up at the Australian Senate, which invited them in writing but cannot compel them, and in Washington, it remains to be seen when the FTC requests arrive and if they reach more labs than the three mentioned.