Why it mattersA coding agent that rewrites its own system prompt every week is hard to script against. Teams need to plan for that churn.
A post titled "Goodbye Claude, Hello Pi" reached the Hacker News front page on 7 November 2026 with 404 points. The author publishes as Caleb at caleb.software. He explains why he stopped using Anthropic's Claude Code as his main coding agent. He has switched to Pi, a smaller open coding agent.
Caleb writes that Anthropic has released Claude Code 244 times this year. He says that over three quarters of those releases changed the system prompt or the tool descriptions. He treats the number as a working-day problem. It is not a change log for him. A tool whose prompt moves every week is hard to build scripts and evals around.
The token count is the sharp end
He spent the most time on the system prompt. Claude Code's Bash tool description takes "almost 3,000 tokens, most of which is random stuff," he writes. Every call to the model pays for those tokens. Anthropic recently cut about 80 percent of the system prompt. Caleb says Claude Code still carries "way more overhead than Pi".
He runs the math for a large user. At 100 million requests per day, he says, trimming 1,000 tokens off could save millions of dollars per year. That is his estimate. It is not Anthropic's figure. He labels it as such.
One specific bug, open since December 2025
Caleb lists other complaints as well. He names a VSCode integration bug. It lets only one session connect to the IDE at a time. He says it has been open since December 2025. He writes that he filed and reproduced it, and that it still holds.
Why Pi
Pi is a coding agent he can shape. He writes that he built the Pi extension he wanted "in like 20 minutes". He finds it more stable and more minimal. The system prompt is small enough that he can read all of it. He is not claiming Pi is a better model. He is claiming it is a smaller surface to work against.
The Hacker News thread picked up the system-prompt point hardest. Commenters with large Claude Code bills said they had seen the same token counts. A few said their eval suites break on each release.
A person building a product on a vendor's coding agent is reading one claim here. The vendor is running the agent faster than a team can keep up with. That is a real cost even if the agent is good.
Source
Goodbye Claude, Hello Pi by Caleb. Discussion on Hacker News.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.