I cancelled my Claude Code Max subscription today. I still intend to use Claude models and I’ve enjoyed Claude Code a lot. However, I’ve had an increasingly difficult experience over the past few months and I’ve reached the point where I want to branch out and try something else.

My AI Use Cases

I am not a software engineer, but I’m trying to work more like one. My primary interactions with AI and agents are in a terminal (Ghostty) or IDE (VS Code). I have integrations with Google Workspace, Notion, and Github. I use UI’s for email editing and similar tasks.

My typical daily uses for AI are:

  • Coding
  • Research
  • Drafting sensitive emails
  • Hardware and software fixes and debugging

What I Want

Stable, reliable, and customizable harness(es): My goal is not to become the most powerful and up to date super user of a harness. My goal is to get reliable and predictable results using a harness to support my work in real life. To that end, I want:

  • Features and changes to be stable, reliable, and tested.
  • Standard conventions and practices for harnesses so I can migrate from one harness to another.
  • Documentation that my agents and I can read and follow.
  • To customize my harness(es) for my specific work cases and preferences. When my harness changes, I want it to be because of something that I explicitly did, not a silent change in the background.
  • To experiment with a variety of models.
  • To keep and use models that may be out of date but work well for me.
  • To switch to different providers during outages and not skip a beat.
  • To use local models offline.

Why I’m Cancelling

In short, working with Claude no longer feels like the magical experience it did 3 months ago. It is still an exciting tool, but I do not consider Claude stable or reliable right now. I frequently find myself working around it or in direct conflict with it and I am ready to try alternatives.

Rate and usage limits: From November 2025 - February 2026, after I upgraded to a Claude Max Plan (5x), I only ran into usage limits once or twice. It was more than enough for me to keep an agent running on coding tasks and work on ad hoc tasks as they came up. In the past month, I’ve run into usage limits multiple times per week. I’ve adapted by trying to downgrade from Opus to Sonnet for most tasks and to divide tasks and labor across models to maximize token efficiency. I try to remember to work on side projects during off peak hours. However, this is a lot of handholding and mental energy on my part that is closer to what I would expect from managing APIs for different models. For me, a Claude Max subscription is no longer set and forget.

Outages. Service outages have also become more frequent. It’s a minor annoyance, but each outage reminds me of the risk of relying on one provider for all of my tokens.

Configuration and default changes. Anthropic recently offered $100 in API credits to current Claude Max subscribers. I accepted the credits, but did not realize that in doing so, I also now allowed token usage above my Max Plan usage limit. After a token heavy session, I realized I had accidentally spent $70 of the credits on a fun but low value side project. I put the cap back on, but I found it concerning that I could (a) spend so much on tokens without realizing it and (b) unknowingly change my spending limits and defaults.

Anthropic is building and growing incredibly quickly. I don’t expect perfection. I also don’t expect Anthropic to subsidize tokens forever. However, the cumulative effect of changes to the user experience, outages, tweaks to the pricing model, token consumption, and overall quality of life degradation have me looking elsewhere.

Harnesses I Want to Try Next

  • OpenCode: I’ve already started using OpenCode a bit and it feels familiar coming from the Claude Code CLI. I think this will be the easiest off ramp from Claude for most tasks.
  • Pi: Recommended to me by a friend. Pi is a minimalist harness. I’m eager to try a streamlined experience with less junk between me and the model. This is the harness I am most excited to try.
  • Cursor: I stopped using Cursor about a year ago, but I’ve heard great things recently. VS Code still feels awkward to use with agents and I don’t like writing/editing code in a terminal.
  • Codex: If I can’t get the results I want with the above, I will try Codex. My primary concern with Codex is recreating the same problems I have with Claude where I’m overly reliant on one proprietary harness and model provider.

Would I come back to Claude?

Absolutely. If I can’t find anything better, I’ll come back. I’ll continue using Anthropic models, just through the API. I don’t feel the magic I felt with Claude in December 2025 - February 2026 and I’m ready to try something new.