"Me? I'll never pay more than €200 a month for AI"

“Me? I’ll never pay more than €200 a month for AI.”

I told myself that for a while. Tomorrow, Fable 5 is coming back. And during the 3 days it was available, I caught something I’d been missing until now.

We’re watching the wrong dial. We ask ourselves “how much am I willing to pay per month”. The real question is: “how big a task can I delegate”.

What changes with this intelligence isn’t the quality of a 2-minute one-shot. “Make me the slides”, “code me this feature”, we already know how to do that. What changes is the horizon. The length of the task an agent can hold on its own.

You go from a question settled in 2 minutes to a full, multi-phase workflow. Its nickname at Anthropic: “dynamic workflows”. Something that used to take days, compressed into a few hours.

The kind of thing that’s coming: you fire off ONE prompt, and the same agent chains everything on its own for an hour.

1. It mines customer feedback on a business problem, stored in ProductBoard, via MCP.
2. It crunches the codebase to understand what exists and its limits.
3. It launches 3 sub-agents in parallel, each with a different solution track.
4. It has them red-team each other, keeps the strongest, iterates once or twice.
5. It opens the Linear ticket with the functional design.
6. It pings you on Slack to validate.
7. You say yes: it tackles the technical design, codes, reviews and opens the MR.

All without you kicking off anything again. A single workflow, wired into your tools.

As devs, we get access to this early. The lucky few. But it’s not a dev thing, it’s going to hit every job.

“Cool, it runs for 1h, you’ve got time to grab a coffee.”

No. If we weren’t trying to max out output, we’d know by now. So instead, you launch three in parallel. And you blow through your token budget even faster. And you tell yourself “maybe I’ll take the €500 plan”.

And that’s where the math flips.

A qualified profile at €2,500 net, taxes included, costs the company around €4,500/month. A €500 license is barely more than 10% of that cost. The agent only has to make them 10 to 20% more productive to pay for itself.

So companies will say yes. From a purely economic standpoint, it’s rational. Ecologically, clearly not…

Meanwhile, we’ve already had two heatwaves and it’s only July 1st. But I see zero path where the appetite for tokens goes back down. I see the opposite: people building training programs to run their agents 24/7, teams measured on tokens consumed per employee, and providers preparing datacenters in orbit because they’ve realized it won’t fit on Earth.

Still, big up to Sonnet 5, which is plenty for every 2-minute one-shot.

Tomorrow, when you test it, don’t just look at whether it codes better. Look at how far you can let it run alone.