Until just two weeks ago, I was a heavy Claude user. I relied on it daily at work and even maintained a personal subscription for side projects, hacks, and experiments.
A subscription to a single AI provider creates a comfortable illusion: that your preferred workflow—sticking with the “best” model for every task—will last indefinitely, right up until you inevitably hit daily or weekly limits.
But that model isn’t sustainable.
The current tokens economy is heavily subsidized. Prices this low can’t last forever. These deals are designed to build deep dependency on one provider. When circumstances change—whether due to pricing shifts, rate limits, policy updates, model deprecations, or strategic decisions—you’re forced to adapt. Suddenly, your carefully honed prompts, agent workflows, context strategies, and muscle memory no longer work as well (or at all).
In the not-too-distant future, the winning approach will likely be intelligent, mixed setups: intelligently routing between frontier cloud models and capable local/open-source LLMs depending on the task, cost, privacy, and speed requirements. That day isn’t fully here yet, but we can start building the flexibility now.
With that in mind, I decided to make the switch proactively. I canceled my Anthropic subscription and moved to a pay-as-you-go setup using OpenRouter + OpenCode. No monthly fees—just transparent consumption-based pricing.
What is OpenRouter and why does it matter?
OpenRouter is a unified API marketplace that gives you access to 400+ models from dozens of providers (Anthropic, OpenAI, Google, Z.AI, Grok, together with many open-source and specialized models) through a single OpenAI-compatible endpoint. One API key, one base URL, and you can switch models instantly—no code changes required.
It matters because it breaks vendor lock-in. You stay in control of costs, can experiment with the best model for each specific task (reasoning, speed, coding, creative work, etc.), benefit from competitive routing and pricing, and future-proof your workflows. It’s the antidote to the “all eggs in one subsidized basket” problem we’re seeing play out across the industry.
One of my favorite features is appending :floor to any model slug (e.g., deepseek/deepseek-v4-flash:floor). This tells OpenRouter to automatically route the request to the cheapest available provider for that model. It’s a simple, zero-config way to optimize purely for cost while keeping the exact same model capabilities. Perfect for high-volume or experimental work where performance is good enough and you want to minimize spend. (There’s also :nitro for prioritizing speed/throughput.)
Here’s how you can configure it in OpenCode:
{
"$schema": "https://opencode.ai/config.json",
"model": "openrouter/deepseek/deepseek-v4-flash:floor",
"provider": {
"openrouter": {
"models": {
"deepseek/deepseek-v4-flash:floor": {
"name": "DeepSeek V4 Flash (floor)"
},
"moonshotai/kimi-k3:floor": {
"name": "Kimi K3 (floor)"
},
"anthropic/claude-sonnet-4.6:floor": {
"name": "Claude Sonnet 4.6 (floor)"
},
"zhipu/glm-5.0:floor": {
"name": "GLM 5.0 (floor)"
},
"zhipu/glm-5.2:floor": {
"name": "GLM 5.2 (floor)"
}
},
"name": "OpenRouter"
}
}
}From there, you get a flexible, terminal-first agent that can tap into whatever model makes sense for the task—without subscriptions or lock-in. I’m already finding it more adaptable and cost-effective for my personal work.
This setup also naturally incentivizes workflow optimization. When you’re not paying a flat monthly fee, every prompt, context window, and tool call has a real cost attached. You start thinking critically about whether you need a frontier model for a given task or if a cheaper, faster model will do. Over time, this pushes you toward more efficient workflows, striking the right balance between quality and cost for each specific use case rather than defaulting to the most expensive option for everything.
I’m still early in this transition, but the mindset shift feels right: treat AI tools as interchangeable infrastructure rather than a permanent home.