noticias ia para empresas agentes de ia claude

AI for automation is getting cheaper and changing without warning: Anthropic has cut the cost of its small model by about 75% with Claude Haiku 5.5, and OpenAI is moving paid ChatGPT, business plans included, to GPT-6. Agents are moving onto every employee's computer, and Windows now lets companies limit which files and networks they can reach. The money is flowing to those who measure whether they obey—Arena raised $200 million and is now valued at around $3.1 billion—and to Manus, which raised more than $500 million after deleting some users' data. Picks up from Thursday's radar, October 8.

Claude Haiku 5.5 costs 75% less: what you didn't automate because it was too expensive now pays off, but measure it by task

On Wednesday, Anthropic launched Claude Haiku 5.5, its small model built for repetitive, high-volume work: summaries, sorting requests, database queries and live customer support. It costs $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens, 90% less than Haiku 4.5; since the new model uses slightly more tokens per job, Anthropic puts the real average savings at around 75%. It also halved the price of Sonnet 5.5 cache reads, making it about 20% cheaper on most agentic work, and is giving subscribers monthly API credits: up to $500 a month on Team plans, pooled across users. HubSpot says it's the best score it has seen on its CRM test suite (92.8%). For your company: if you kept sorting emails or support tickets by hand because AI was too expensive, that excuse just expired. But compare cost per finished task, not per token: as we reported on October 6, the cheap model can end up costing more. Run 20 real cases through Haiku 5.5 and through what you use today, and log cost and accuracy. If you pay for Claude Team, spend this month's credits on one specific pilot, not on random experiments. And for complex work, Anthropic itself recommends sticking with Sonnet or Opus. A task still done by hand "because AI was expensive" is stuck in MANUAL on math that's no longer true. Source

ChatGPT moves to GPT-6 on business plans: this week, retest what your team asks it every day

OpenAI began rolling out GPT-6 in ChatGPT for Plus, Pro, Business, and Enterprise plans on Wednesday, and for free plans on Thursday: it reaches the more than 1.2 billion people who use it every week. Paid plans get GPT-6 Sol, and with it comes Intelligent UI: responses can include interactive charts, buttons, forms, and small tools built on the fly, like a calculator. Nobody has to change a setting to get it; on Enterprise, it depends on how admins configure it, and only the Chat tab changes: the models behind Work and Codex stay the same. OpenAI admits the model's design judgment still needs work. For your company: the assistant your staff uses has changed models overnight. This week, run your team's five most common requests through GPT-6—the meeting summary, the standard reply to a customer, the Monday report—and compare the results with what you got before. If you run an Enterprise workspace, you decide when it goes live. And one rule for the new feature: a calculator ChatGPT whips up in a chat isn't a validated tool; if one of its numbers ends up in a quote, someone checks it. A process that switches models without anyone testing it isn't INTEGRATED: it runs on someone else's schedule. Source

Windows now lets you set which files and networks each AI agent can touch: decide before Copilot acts on your team's laptops

On Wednesday, Microsoft laid out its plan for agents to work on the PC itself, not just in the cloud. The piece that matters for businesses is already here: Microsoft Execution Containers (MXC), in Windows 11, allows IT to define which files and network destinations an agent can use, and enforces that while the agent runs, outside its control: the agent can't grant itself more access. OpenAI's Codex, GitHub Copilot, OpenClaw and Replit already support it, and Claude Code, Manus and Perplexity are on the way. Copilot will also be able to use files from the PC, act within Windows—sorting files, diagnosing the computer—and use local models, always with user permission, on Copilot+ PCs "in the coming months"; Microsoft says they already make up more than 40% of laptops being built for business. Soon, Windows will distinguish what the agent did from what the person did. For your company: agents will arrive through each employee's computer, not through an IT project. Take stock now of the ones already installed and, if you have an in-house IT person or an IT provider, ask for a policy for each agent: which folder it can touch, which websites it can reach and what stays off-limits, like personal documents or the accounts. It's the same principle as onboarding an agent like an employee. An agent working with all its user's permissions is not UNATTENDED: it has the keys to the whole house. Source

Arena, the most-watched AI model leaderboard, raises $200 million to measure which agents do things nobody asked for

Arena, the company behind the industry's most-watched AI model leaderboard, has raised $200 million at a valuation of approximately $3.1 billion, nearly double its January valuation of $1.7 billion, according to Bloomberg. Lightspeed and Khosla Ventures led the round, with Salesforce Ventures and Dell Technologies Capital joining. Alongside the round, it's launching an Alignment Index that scores more than two dozen models on how often they act without authorization and how often they claim to have finished work they didn't; GPT-6.1 Sol tops it and Claude Opus 5.5 comes second. In Europe, Galtea, a Barcelona Supercomputing Center spin-off, does something similar: it raised $3.2 million in March to run adversarial tests on agents and hand the results over as evidence under the EU AI Act. For your company: the money is going to whoever measures whether agents do as they're told, because that's the risk holding companies back. Use the index as a first filter, not as a guarantee: what counts is checking, on your own tasks, whether the agent says "done" when it isn't. Before you let one run on its own, review its first 20 tasks by hand and note how often it went beyond what you asked or marked something finished that wasn't. An agent whose "done" nobody checks isn't UNATTENDED: it's flying blind. Source

Manus, the agent Meta had to give back, raises more than $500 million after deleting some users' data: make sure you can export yours

On Thursday, Butterfly Effect, the company behind the Manus agent, announced a round of more than $500 million led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund. It's the company's first raise since China forced Meta in April to unwind the $2 billion acquisition it had announced in December; it didn't disclose a valuation, though a $4 billion figure was reported in September. The split came at a cost for users: to comply with regulators, in late August Manus deleted data that certain accounts had generated since December 29, after giving them a window to back it up and restore it later. For your company: your agent provider can change hands—or be ordered by a government to undo the deal—and your data goes along for the ride. Before you trust any agent with real work, check three things: whether you can export what it produces in a format you can open without it, where the data is processed, and what the contract says if ownership changes. And keep your own copy of what you cannot afford to lose. A process whose only file lives in someone else's agent is not INTEGRATED: it's on loan. Source

What to watch tomorrow

Whether Anthropic switches on the monthly API credits for Team plans it promised for this week, and how long ChatGPT Enterprise admins take to open GPT-6 to their staff. And on Tuesday, October 13, TechCrunch Disrupt kicks off in San Francisco, running through the 15th.