🔄 Update: in the end I did not move the agents to ChatGPT – they stayed on Claude and only switched from the Agent SDK to the Claude CLI. How that went is in the post "Moved from Claude SDK to CLI"
⚡️ The weekend is already booked: I'm moving 14 agents from Claude to ChatGPT. On May 13 Anthropic announced - starting June 15, Agent SDK and claude -p move out of the subscription into a separate credit pool
What exactly changes on June 15:
Anthropic is separating agent workloads from interactive use. Until June 15 - as many agent calls as you want within your subscription. After that - a credit pool by plan: Pro $20/mo, Max 5x $100/mo, Max 20x $200/mo. Credits expire every month, they don't roll over. This moves agent traffic off subscription pricing onto a rate closer to the API
On the meter: Agent SDK, claude -p, Claude Code in GitHub Actions, third-party apps with Agent SDK auth. Not affected: interactive Claude Code in the terminal and the web chat. If you work through an API Key, you won't notice any change. The ones who get hit are specifically developers who ran automation on a consumer subscription
Eligible users will get an email with instructions to claim the credit before June 15 - check your inbox and your subscription settings
Why 14 agents is a problem:
📌 All 14 live on a VPS - running in the background in Telegram, handling incoming messages, routing requests
📌 All of them go through Agent SDK - exactly what Anthropic is putting on the meter
📌 14 parallel agents - that's not "a few rare calls a day"
Starting June 15 all this traffic will eat into the Pro credit. $20 a month for 14 parallel agents is a hard limit that'll run out fast
ChatGPT Pro still works without a separate agent meter and includes Agent Mode, GPT-5.5 and Deep Research. Right now ChatGPT is just as good as Claude on agentic tasks - and with an almost unlimited subscription it's the obvious choice for background automation
My plan for this weekend:
🟢 Moving to ChatGPT - all 14 Telegram agents on the VPS
🟢 Staying on Claude - interactive Claude Code in the terminal, all the hands-on coding
🟢 Deadline - reconfigure, test and switch the traffic before June 15
How it ended
I now run 8 Claude Code agents every day, all on a subscription, with no separate bill for every request. The move taught me the main thing: a provider's pricing can change at any moment, and the system has to survive that without panic. That's why my personal agent Jes has a primary model and an automatic backup: if the first fails or hits a limit, the second carries on by itself. Jes has been running for 134 days and handles 10 scheduled jobs on its own, and no change of terms has stopped it in that time
How I build systems for clients now
For clients I set the system up on their own server and their own AI subscription. That's how AI assistants run for five people, and how a clinic owner runs 6 agents. The bill for the models goes straight to the provider, so if the terms change, the client sees it at once rather than through my markup. The model inside the system can be replaced without touching memory, rules or the schedule: those are stored separately from it
Advice for the future: never keep a business process on one model and one plan without a fallback. Check in advance what happens if your provider introduces a meter or raises the price tomorrow. If the answer is "everything stops", that's your weakest point
Simple arithmetic explains why news like this triggers moves. My own spending on models and a server is now about $300 a month. If I paid for the same volume of work per request through an API, by my estimate the bill would be roughly a hundred times larger. For agents that run around the clock, per-request billing changes the whole economics. But count your own volume first, then move
🙃 This is how it's going to work now. The free ride had to end sooner or later, and better sooner 😎