A Thursday with a new model from Google, a hidden cost in Claude Code and agents that found a way to talk without a network.

Anyone who works in product has been through this: the new model ships, the price looks great, and only later do you read the fine print.

Yesterday Google announced Gemini 4 Carbon, focused on cybersecurity and long engineering tasks. According to the company, it ranked first on the Artificial Analysis index, with 59 points, and scored 88.2% on SEC-bench Pro, which measures the ability to find vulnerabilities. These are Google’s own numbers, so it is worth waiting for independent tests.

The price that doubles

What caught my attention was the price. The introductory price is US$ 2 per million input tokens and US$ 10 for output. After the launch period, it goes up to US$ 4 and US$ 20. In other words, whoever builds the math for a feature on today’s price will see the cost double later on. The guide on AI agent cost management and the cost budget template help build that math at the full price.

Then there is access. At first, only a restricted group of cybersecurity defenders uses the model, and for the rest of developers and companies Google says “as soon as possible”, with no date.

What changes for the product

On the product side, the lesson is practical. A roadmap should not depend on a model your team cannot use yet, and the feature’s margin has to close at the full price, not the launch price. This is part of AI product management and of AI agent operations.

Even so, I think it is great to see labs competing on long tasks, with less cheating on tests and more attention to security. That opens room for more useful agents inside products, and it is good for those who build.

For those who got curious and want to read the full announcement, here is the link.

The rest of the radar

  • Claude Code and the hidden cost — a user measured that the call that only labels the status of each background agent accounted for 16.4% of estimated spend and 19.7% of requests; these are Claude Code’s own estimates, not invoices, and there is no official option to turn it off. Gist.
  • Cursor Agent SDK — agents defined in code, with evals, subagents and PR triggers, which the team can version and measure; the page does not state pricing. Cursor.
  • OpenAI agents and the CPU cache — in tests, isolated agents communicated without a network through hidden channels; isolating the network is not enough, and egress control and red-teaming become a product requirement. The Decoder.
  • AI executives and the general public — a TechCrunch opinion piece: 50% of US adults think AI will reduce people’s ability to form meaningful relationships and only 5% think it will make it easier (Pew, Sep/2025); discovery with real users avoids building what nobody asked for. TechCrunch.

Tomorrow there is more, and the idea stays the same: separate what changes the product decision from what is just noise.