Beginning April 4th at 3PM ET, Anthropic is flipping a switch. Claude subscription limits (Pro, Max) no longer cover third-party harnesses like OpenClaw. If you were running Claude through OpenClaw on your subscription, your agent usage will now be billed separately through Anthropic's extra usage bundles, whether you expect it or not.
If OpenClaw was the main reason you were paying $20-200/mo for a Claude subscription, that subscription just stopped earning its keep. It only covers Anthropic's own surfaces now: claude.ai, the mobile app, and Claude Code. If you don't use any of those, you're lighting money on fire.
With that said, I decided it was probably best to migrate to a platform with some redundancy built in, especially given Anthropic's outages in the past. I moved my OpenClaw setup over to OpenRouter and it only took me a few minutes. Right now I'm running a free model as my primary with Claude as an automatic fallback, so my day-to-day usage costs me basically nothing. I'll walk through how I set that up and why it turned out to be an upgrade.
What actually changed
Boris Cherny, one of the execs behind Claude Code, posted the announcement on X Friday evening, April 3rd. The change went live the next day. Your Claude subscription (Pro at $20/mo, Max at $100-200/mo) used to cover API calls from third-party tools like OpenClaw and other agent harnesses. Now it doesn't.
I get it, it's a business decision. Anthropic's subscriptions weren't built for the usage patterns that tools like OpenClaw generate, and they're prioritizing capacity for their own products. Fair enough. But if you're someone who relies on OpenClaw to manage your calendar, automate your morning briefing, or keep tabs on your RSS feeds, this change puts you in a tough spot.
The timing stung. A tweet Friday night, a cutoff Saturday afternoon. According to The Verge, OpenClaw's creator Peter Steinberger (now at OpenAI) and board member Dave Morin tried to get Anthropic to delay, and only managed to buy a week.
Why OpenRouter instead of a direct Anthropic API key
You could swap to a direct Anthropic API key and call it a day. But after going through this, I realized the real issue isn't billing. It's that my entire agent setup was dependent on one provider's decisions. If you're running tasks that actually matter to your day, you deserve something more stable than that.
I've been using OpenRouter across several of my projects, including AccessHawk. It's a unified API that sits in front of dozens of model providers. One API key, one base URL, and you can hit Claude, GPT, Gemini, or whatever else you need without managing separate credentials for each one.
Distributed infrastructure with automatic failover
Anthropic's API has outages. Every provider does. When you're running an agent that monitors your calendar every 15 minutes, checks RSS feeds three times a day, and handles your morning briefing, downtime isn't just annoying. It's missed alerts.
OpenRouter maintains connections to multiple endpoints for the same model. If one Anthropic endpoint is slow or down, it routes to another. If Anthropic is entirely unreachable, you can configure fallback models from other providers. Your agent keeps running while you sleep.
Model freedom without reconfiguration
With a direct Anthropic API key, you're locked into Claude. Want to try GPT-5 for a specific task? New API key, new provider config, new auth setup. Want to use a cheaper model for simple cron jobs? Still Anthropic's pricing tiers.
With OpenRouter, switching models is a one-line config change. Your API key doesn't change. Your auth doesn't change. You just swap the model string:
# Switch from Anthropic's premium tier to mid tier
openclaw models set openrouter/anthropic/<current-sonnet>
# Try OpenAI's flagship for a while
openclaw models set openrouter/openai/<current-gpt-flagship>
# Go back to Anthropic's premium tier
openclaw models set openrouter/anthropic/<current-opus>Substitute the angle-bracketed placeholders with current model IDs from openrouter.ai/models. As of April 2026 those would be claude-sonnet-4-6, gpt-5-turbo, and claude-opus-4-7 respectively.
If you don't want to pick a model at all, OpenRouter has an Auto Router (openrouter/auto) that analyzes your prompt and picks the best model for the task. Simple prompts get routed to cheaper models, complex ones get the heavy hitters. No extra cost on top of whatever model it selects. You can also see which models other OpenClaw users are running on OpenRouter to get a sense of what works.
Cost transparency
OpenRouter shows you what every model costs per million tokens, side by side. When you're deciding whether your daily reminder cron needs Opus ($5/$25 per M tokens) or whether Haiku ($1/$5) would do the job just fine, that visibility matters. It also means you can spot pricing changes immediately instead of getting surprised on your next invoice. (All per-million-token figures in this post are accurate as of April 2026; check OpenRouter's model page for current rates before you wire anything up.)
Insurance against the next billing change
This is the one that sold me. Anthropic changed their billing with 18 hours notice. If you're on a direct API key and they change pricing, rate limits, or terms, you're scrambling. With OpenRouter, your worst case is changing a model string. Your agent infrastructure, auth, and configuration stay untouched.
The migration: 5 minutes, 3 steps
If you made it this far, it may be obvious that this assumes you already have OpenClaw installed and running. If you're starting from scratch, check the OpenClaw getting started guide first. Otherwise, the migration from Anthropic subscription auth to OpenRouter looks like this.
Step 1: Get an OpenRouter API key
Go to openrouter.ai/keys, create an account, add credits (start with $20-30), and generate an API key. It starts with sk-or-.
Step 2: Add OpenRouter to your OpenClaw config
Open your openclaw.json and add the OpenRouter provider. Set your API key and your preferred default model:
{
"agents": {
"defaults": {
"model": {
"primary": "openrouter/anthropic/<current-opus>"
},
"models": {
"openrouter/anthropic/<current-opus>": {}
}
}
},
"env": {
"OPENROUTER_API_KEY": "sk-or-..."
}
}Replace <current-opus> with whatever's current on OpenRouter (e.g., claude-opus-4-7 as of April 2026, or openrouter/auto if you'd rather have OpenRouter pick a model per prompt).
This doesn't remove your existing Anthropic config. You can always switch back. For the full configuration reference, see the OpenClaw OpenRouter provider docs.
Step 3: Restart the gateway
# Restart the gateway to pick up the new config
openclaw gateway restart
# Verify the model is active
openclaw models statusThat's it. Your next conversation will route through OpenRouter to Anthropic's Claude Opus. Same model, same quality.
Optimizing costs with per-cron model selection
This is where OpenRouter really pays off. Not every task needs your most expensive model. A simple daily reminder doesn't need Opus-level intelligence. Neither does a git auto-commit script or a simple email check.
OpenClaw lets you set a model per cron job. Combine that with OpenRouter's access to every pricing tier, and you can optimize aggressively:
# List your cron jobs to get the job IDs
openclaw cron list
# Lightweight crons: use the mid tier (cheaper per token)
openclaw cron edit <job-id> --model openrouter/anthropic/<current-sonnet>
# Complex crons: keep the premium tier
openclaw cron edit <job-id> --model openrouter/anthropic/<current-opus>
# Restart the gateway to pick up the changes
openclaw gateway restartAfter editing, you can verify which model each cron is using:
openclaw cron list --json | jq '.jobs[] | {name, model: .payload.model}'The pricing difference is real. Sonnet's output tokens cost 40% less than Opus. For crons that run every 15-30 minutes and only need to parse some text or run a shell script, that adds up fast.
A typical setup with 5-7 cron jobs and regular conversation might look like this:
- Conversations (Opus): $20-40/month depending on how chatty you are
- Complex crons (Opus): $5-10/month for 2-3 jobs with multi-step tool orchestration
- Simple crons (Sonnet/Haiku): $3-8/month for 4-5 lightweight scheduled tasks
- Total: roughly $30-60/month, often less than a Max subscription
With prompt caching enabled (which works through OpenRouter on Anthropic models), cached input tokens cost 90% less. If your crons use similar system prompts across runs, caching alone can cut your bill significantly.
My current setup
I'm currently using Qwen 3.6 Plus as my primary model while it's free on OpenRouter, with the Auto Router as a fallback. I've configured the Auto Router's Allowed Models to only select from Anthropic models, so if the free model goes down or gets retired, my agents fall back to Claude automatically.
{
"agents": {
"defaults": {
"model": {
"primary": "openrouter/<current-free-or-cheap-model>",
"fallbacks": ["openrouter/openrouter/auto"]
}
}
}
}In April 2026 my primary was openrouter/qwen/qwen3.6-plus:free; pick whatever's free or cheap on OpenRouter when you read this. For the time being, this costs me basically nothing for day-to-day usage. When the free tier goes away, the Auto Router kicks in and picks the best Anthropic model for each prompt. That kind of setup just isn't possible when you're locked into a single provider.
What about the other options?
OpenRouter isn't the only path forward. I looked at the alternatives before committing.
Direct Anthropic API key
Straightforward to set up with openclaw onboard --anthropic-api-key "$KEY", and the pricing is the same as going through OpenRouter for Claude models. But you're still tied to a single provider with no failover and no model flexibility. If that's fine for your use case, it works. I just didn't want to be in this same position the next time something changes.
Claude Desktop with Cowork scheduled tasks
Claude Desktop's Cowork feature now supports scheduled tasks with /schedule. Sounds like a replacement for OpenClaw crons, but there are significant gaps. Your Mac has to be awake and the app running; if it sleeps, tasks get skipped. There's no persistent memory system, no heartbeat coordination, and no way to send results to Telegram or Discord natively. Each task runs as an isolated session with no shared context. For a simple "summarize this document every Monday" workflow, it's fine. For an agent setup with interconnected crons, memory, and mobile notifications, it's not there yet.
Switch to a different provider entirely
OpenClaw supports OpenAI, Google Gemini (API key or OAuth through the Gemini CLI), Qwen, GLM, MiniMax, and others. Some offer flat-rate coding plans. If you're open to switching away from Claude entirely, you could use your Google One AI Pro subscription for Gemini access at no extra cost. The tradeoff is that Claude's tool use and instruction following is still best-in-class for complex agent workflows. Gemini and GPT are catching up, but for multi-step tool orchestration with nuanced system prompts, Opus is hard to beat.
The bigger takeaway: don't depend on a single provider
This billing change was a reminder of something we already know in infrastructure: single points of failure are liabilities. Your database has replicas. Your deployment has rollback. Your DNS has failover. Your AI provider should too.
OpenRouter isn't the only routing layer out there, but it's the one with the best OpenClaw integration today. The point isn't "use OpenRouter." The point is: abstract your provider dependency so that the next time a vendor changes terms on a Friday night, you're not spending your weekend reconfiguring your entire agent setup.
I run cron jobs that check my calendar, monitor RSS feeds, send me security advisories, and automate git commits. None of those should break because a billing department made a policy change. With OpenRouter in the middle, they didn't.
Conclusion
The migration took five minutes. An API key, a config edit, a gateway restart, and everything was back. My crons kept running, my Telegram conversations kept working, and nothing broke.
If you're an OpenClaw user still running on Anthropic subscription auth, the move is simple: get an OpenRouter API key, add it to your openclaw.json, and restart the gateway. You can run free models day-to-day with Claude as an automatic fallback, or just use Claude directly through OpenRouter with failover built in. Either way, you're not at the mercy of one provider's billing decisions anymore.
The OpenClaw docs have a full OpenRouter provider guide if you want the detailed configuration reference.