I've been going back and forth between Claude Code and OpenCode since last year. I started with Claude Code and the Max subscription for client work with Drupal, then moved to OpenCode because it let me plug in cheaper models, and when Anthropic cut off using the subscription through OpenCode I went back to Claude Code for the heavy lifting. Since then I've kept a foot in each camp: Claude Code for interactive work, OpenCode for the repetitive tasks with cheap models. This week, after seeing what Anthropic's done, I'm pretty clear on where I'm shifting the weight.
They did two things three days apart. On June 9 they released Fable 5, their most powerful model to date, a new class they themselves place above Opus. And for the 15th they announced that claude -p, the headless mode I use to automate, stops counting toward the subscription. On their own, they look like two unrelated bits of news. Together they tell one single story, and I don't like it.
The good model, on a two-week trial
Fable 5 is included in the Pro and Max plans. But only from June 9 to the 22nd. On the 23rd they pull it, and after that using it costs extra, at API prices: double what Opus costs. Anthropic says they'll bring it back to the subscription "when capacity allows," with no date attached.
So the smartest model they have isn't included in what you pay for. It's a free two-week trial and then you go through the till. And something tells me this is going to be the new normal: the newest, more capable models stay outside the plan, and inside you get the usual stuff a few months late.
What really gets me: the credit that evaporates
The claude -p thing goes the same way. From the 15th, headless mode, the Agent SDK and the GitHub Actions integration stop drawing on your plan's limits. They move to a separate monthly credit: 20 dollars on Pro, 100 on Max 5x, 200 on Max 20x. What stays inside the subscription is interactive use, meaning Claude Code in the terminal or IDE, the chat, and Cowork.
Fine so far. But there's a detail in the small print that's the one that got to me: that credit doesn't roll over. If you don't spend it in a month, you lose it. And when it runs out, either you've got usage credits enabled and keep paying at API rates, or your SDK requests just stop until next month.
The subsidy is winding down
What was happening until now is that the subscription was subsidizing programmatic use. This isn't something I'm making up. The Zed folks worked out that the plan came out between 15 and 30 times cheaper than the API for that kind of use. It was a sweet deal, we all knew it, and that's why half the community built their loops and their agents on top of the subscription.
Well, the sweet deal's over. And what bothers me isn't that it's ending, which I get, but how they're doing it. Instead of raising the price and saying so to your face, they hollow out the plan from the inside. You pay the same and every month there's less in it. The good model gone, automation gone. It's the third time this year they've tightened the screw, and the pattern always points the same way. They change their mind as they go and take away things you used to have for the same money. I find it pretty ugly, honestly.
Let me be clear: I'm not bashing them for free. Running frontier inference costs a fortune and somebody has to pay for it. I get that completely. But there's a difference between raising the price openly and keeping it the same while cutting what you offer. The first I respect. The second I don't.
I've already got the alternative set up
This is where I'm glad I never threw away the OpenCode setup. I've got my ddev-ai-workspace, which I've already written about on the blog, and the nice thing is it puts OpenCode and Claude Code in separate DDEV containers inside the same project. Switching from one to the other is literally typing ddev oc instead of ddev cc. Both share the same agents, the same Playwright, the same task tracking. And Ralph, the autonomous orchestrator, you tell it --backend opencode or --backend claude and it runs the same thing underneath.
So the cost of making the jump is basically zero for me. I don't have to migrate anything or set up a new environment: I've been keeping both running in parallel for months, it's how I work today. All I'm going to do is move the work I currently do in Claude Code over to OpenCode.
And in OpenCode the play I'm after is combining models. The trick is to separate thinking from doing. The expensive part of any task is the reasoning: understanding the problem, deciding the architecture, planning the steps. A smart model does that once. What comes after is following the plan, and a cheap model handles that just fine. It's the same subagent logic I already covered for Claude Code, except in OpenCode I pick each piece myself, without depending on whatever Anthropic decides to include that month.
In practice, for the orchestrator I put something good at planning, GPT or Opus itself via API. For the workers, for the bulk of the job, cheap models: DeepSeek V4, which is about as cheap as it gets right now and performs surprisingly close to the frontier, or Kimi K2.6 if I want open weights. And for whatever I don't want leaving the house, Qwen running locally, which on a Qwen3-Coder runs on a regular GPU and costs basically nothing per token. The cheap model does the bulk and the expensive one only steps in for the hard parts, so you save a pile of money without losing quality.
Where Claude still pays off for me
To be fair, I'm not throwing the subscription in the bin entirely. The web chat mode, especially for deep research, is still some of the best out there and I use it a lot. That I am keeping, but the 20-dollar plan is more than enough for it. I don't need a 100 or 200 a month Max plan to do research in the browser.
What no longer adds up is paying for the expensive subscription for agentic and coding tasks. That's where Claude Code makes less and less sense for me and OpenCode more and more. Combining models, keeping the spend under control, and not depending on terms that change every quarter outweighs the convenience of having it all inside one plan.
That said, setting up this kind of thing isn't for everyone. You have to know which model goes where, watch that a cheap one doesn't slip a botched job into production, juggle several API keys, and review what comes out. If you don't fancy getting into that mess, the subscription hands you everything ready-made and for a lot of people it'll still be worth it. But anyone selling you that rolling your own orchestration is plug and play is lying.
What I'm going to do
My plan is simple. The rest of the month, with the subscription already paid, I'm spending on testing properly, with real client work, not toy examples. If the numbers come out halfway decent, and right now it's looking that way, next month I drop the Claude Code subscription and do ninety-something percent of my work in OpenCode alone. The 20 plan I keep only for web chat and deep research.
What I'm really going to measure this month isn't so much the savings, which I've got a pretty clear picture of, but whether the cheap models hold up in real Drupal work. Cutting the cost is one thing, the result staying good is another. If they start slipping nonsense into production, I won't have gained a thing.
But something tells me Anthropic isn't going to change course, and that every quarter the subscription will give a little less for the same money. If that's how it goes, sooner or later you have to look out for yourself. And I'd rather start doing that now, calmly and with both environments already set up, than the day they force me to.
PS: If you want to see how I've got all this wired up, I explained the model-per-subagent split in the article on how to optimise token usage in Claude Code, and the full environment with both CLIs running in parallel is in ddev-ai-workspace, which is free and installs with a single command.