Adieu, Claude
When Anthropic proved to be a liability, I replaced Claude Code with OpenCode.
I already ran Crossfire and a few private projects through OpenRouter on a pay-as-you-go basis. Claude Code was the subscription I did not have to think about for twenty bucks a month, but I no longer trust the company behind it.
Fable’s fairy tale
Anthropic shipped Fable 5, silently rerouting certain requests to the weaker Opus 4.8. The affected requests were related to cybersecurity and anything resembling model distillation. Anthropic admitted it had made the wrong decision once the backlash made it impossible not to.
A few days later, the export-control directive grounded Fable 5, for reasons Anthropic itself called a misunderstanding. I do not much care whose fault that part is. Governments issue directives.
A fairy tale retold
I had already dropped OpenAI before this. They claimed GPT-2 was too dangerous to release, which was rubbish, and the “Open” was gone the moment they smelled stacks of money. Then they put ads in ChatGPT. When Anthropic refused the Pentagon’s demand to remove red lines on fully autonomous weapons, OpenAI stepped in hours later and signed the deal.
Two weeks after Anthropic’s models were pulled, OpenAI announced its latest models and then restricted them to “trusted partners” at the US government’s request. The US has made it impossible for me to treat access to frontier models as a sure thing.
Crossing the drawbridge
The only reason the switch from Claude Code to OpenCode was painless is that I built on open foundations before I needed them, which is why migrating off Logseq was relatively painless, too. The coding agent talks to Zed through the Agent Client Protocol, an open standard. I had built more than a dozen skills for Claude Code, yet OpenCode read them without a single change, because they are plain Markdown with YAML front matter. OpenCode is MIT-licensed and I could run it myself if I wanted to. None of that was true with Anthropic or OpenAI.
Lock-in is the moat, not the data or the model architecture. It is to keep customers from leaving, not necessarily to keep competitors from entering. That moat is what keeps you captive when the company decides, as it inevitably does, to squeeze customers. Open standards and open source are the only guarantee against that betrayal.
The toll
I already knew what pay-as-you-go felt like. At the end of the month I often asked myself whether it was worth it. Claude Code had a flat fee and I never asked that question. OpenCode Go gives me the same at ten dollars a month, and they are honest about what that buys: “generous limits and reliable access to the most capable open source models.” Sign up through that link and we both get five dollars in credit, which is the extent of my commercial interest in this post. Twenty dollars a month for OpenAI or Anthropic is twice the price, and the $100–$200 plans are in a different league entirely. You get what you pay for: a single prompt against GLM-5.2 on a large codebase burns through the quota like an AI startup through $100m, which is where the truly free models fill the gap.
The models are good enough for the work I described. Day to day, I use Big Pickle, a model OpenCode will not name and only calls a stealth model, free for now while the team collects feedback. It handles basic refactors, scripts, and multi-file edits without a fuss. For harder tasks I switch to GLM-5.2, which scored 81 against Claude Opus 4.8’s 85 and GPT-5.5’s 84 ahead of Gemini 3.1 Pro (74), though of course benchmarks are not everything.
I could buy the same GLM models through Z.ai, but its platform separates Chinese and overseas infrastructure and reliability complaints are rife. OpenCode Go runs the same models through servers in the US, EU, and Singapore instead. That distinction is the entire reason I am comfortable using GLM at all.
The moral of the story
None of this makes me independent of frontier labs in any complete sense. The harder tasks still send me back to Zen, OpenCode’s pay-as-you-go option, where Claude and GPT are available—Gemini is there too, but its tendency to roll over like a puppy makes it useless when I need it to disagree with me. The only full exit from that arrangement is running open models on my own hardware, which is fine for autocomplete and short questions though not yet usable for anything that requires the model to “think.” The OpenCode quotas are a bit of a bummer, but money can fix that problem. Trust cannot be bought, and that is the moral of the story.