noticias radar ia precios ia agentes ia ai act

Yesterday's radar closed with agents as an attack and regulatory surface; today the weekend adds the other side: automating with AI has just become much cheaper. OpenAI cut the price of its budget model by 80%, DeepSeek released an open agentic model at a third of the cost, and a startup raised 30 million to clone your best employees into agents; meanwhile, as of yesterday, the Spanish AI Supervision Agency can inspect and fine any company using AI. The weekend reading: it’s cheaper and easier to automate than ever—just as Spain starts having someone watching how you do it.

Automating routine tasks with AI costs five times less today than on Friday

On July 30, OpenAI cut the price of GPT-5.6 Luna, its fast and cheap model for daily work—summaries, writing, classification, automated support—by 80%: from $1 to $0.20 per million input tokens, and from $6 to $1.20 for output. Terra, the mid-tier model, drops by 20%, while the flagship Sol remains unchanged. The cut comes just three weeks after the family's launch, a sign that competitive pressure is eroding the margins of major labs. For your company: if you have workflows that summarize emails, write drafts, tag tickets, or power a chatbot, the variable cost of those processes has just plummeted; review which model your automations use and drop to the budget tier where quality holds up, because most routine work doesn't need the expensive model. That project you discarded a month ago for being too expensive is worth crunching the numbers on again today. Source

The engine behind that drop: an open agentic model at a third of the price

On July 31, DeepSeek updated its V4-Flash model (version 0731) with major improvements in agent and programming tasks, achieved by retraining the existing model without changing its architecture or size; it now outperforms its own Pro version in all published agentic benchmarks. It maintains the price at $0.14 per million input tokens and $0.28 for output—a third of its Pro range—natively speaks the OpenAI Responses API format, is compatible with Codex, and its weights are open under an MIT license. For your company: this is the kind of pressure that forces cuts like the one above, and it is directly exploitable: a model good at agents and code, cheap, and that fits almost as a direct replacement in tools that already talk to OpenAI. If you work with a technical provider, ask them to route agentic or code tasks to a model like this and compare cost and quality before renewing contracts. Source

New agents don't divert calls: they copy your best salesperson

Encore AI raised $30 million in a Series A round on July 29 for a technology it calls "interaction mining": it analyzes your company's calls, chats, emails, and CRM, extracts exactly what your top-selling employees do, and deploys agents that replicate those behaviors across all channels. The shift from the previous wave of agents is the goal: not reducing how many people reach a human, but increasing the revenue from every interaction. For your company: the signal is that the value of an agent is no longer just in the model, but in your own operational data—your best conversations are the asset. You don't need Encore to start: organize and save the recordings and transcripts of your best sales and support reps, because they are the raw material with which you or a provider will train the agent that standardizes what only two or three people know how to do today. Source

In Spain, as of yesterday, someone can now knock on your door because of your AI

With the entry into force of the sanctions regime of the European AI Act, the Spanish AI Supervision Agency (AESIA), based in A Coruña, debuted its formal powers yesterday: it can request information, inspect AI systems, demand corrective measures, and issue sanctions, with fines of up to 35 million euros or 7% of annual turnover in the most serious cases. It's not just Brussels watching the big providers: it's a national body with jurisdiction over any company using AI in Spain, and its sanctions can be compounded with GDPR fines if personal data is involved. For your company: the immediate risk for an SME isn't that 7%—that targets serious non-compliance—but the transparency obligations that do apply to you and now have someone to enforce them: notifying that your chatbot is a bot, labeling content generated by your AI, and knowing how to document what systems you use and for what. Spend an afternoon doing that inventory and putting up notices; it's cheap today and expensive when they ask for it. Source

What to watch this week?

This week: if the price war continues, expect responses from Google and Anthropic to OpenAI's cut; the Ai4 conference (August 4-6) will serve as a thermometer for agent adoption; and in Spain, the first signs of how AESIA debuts its powers. And a question for Monday: now that automating is cheaper and your best employees are trainable, which process are you going to move to the budget model this very week?

Watch the 1-minute video