Anthropic's OpenClaw crackdown: what it means for your API bill

By Linas Valiukas · April 7, 2026

Three days ago, Anthropic cut Claude subscriptions off from OpenClaw. We covered what happened. Now let's talk about what it costs.

Because the sticker shock is real. People who were paying $20/month for Claude Pro are staring at bills that could hit $200, $500, or more. The math isn't obvious, and OpenClaw's architecture has a few quirks that make it way more expensive than you'd expect.

The new rates

With subscriptions out, you're on Anthropic's standard API pricing. Per million tokens:

Those numbers look manageable in isolation. A million tokens is roughly 750,000 words. How bad could it be?

Bad. Because OpenClaw doesn't use tokens the way you think it does.

The heartbeat tax

OpenClaw's heartbeat is a recurring timer that wakes your agent up to check for tasks, run cron jobs, and maintain state. Default interval: 30 minutes. Sounds fine. Here's the problem.

Each heartbeat carries your full session context. Chat history, injected workspace docs, SOUL.md personality file, skill definitions. A typical heartbeat run pushes 170,000 to 210,000 input tokens. Every single time.

On Opus at $5/M input tokens, that's roughly $0.85 to $1.05 per heartbeat. At the default 30-minute interval, you're burning 48 heartbeats per day. That's $41 to $50 per day just from heartbeats. You haven't sent a single message yet.

One user on GitHub documented this. Their automated email check, set to run every 5 minutes, burned $50 in a single day. Another reported $3,600 in a month from what they thought was light usage.

When you were on a flat-rate subscription, none of this mattered. The meter wasn't running. Now it is.

Context bloat: the 40% you don't see

The biggest line item on most OpenClaw API bills isn't messages you send. It's context accumulation - and it accounts for 40 to 50% of total token consumption.

Every time OpenClaw calls the API, it sends the entire conversation history plus system context. As your session grows, so does every subsequent API call. A fresh session might cost $0.02 per message. After an hour of back-and-forth, that same message costs $0.15 because you're re-sending the whole conversation every time.

Anthropic's own products handle this with prompt caching - reusing previously processed context at a 90% discount. Claude Code is optimized for high cache hit rates. OpenClaw isn't. Dynamic system prompts, injected metadata, conversation structures that don't match Anthropic's caching assumptions - OpenClaw sessions typically get around 10% cache hit rates versus Claude Code's much higher numbers.

That's not a rounding error. A 10% cache rate versus an 80% cache rate on the same number of tokens means you're paying roughly 4x more per interaction for the same work.

Real bills from real users

Here's what people are actually reporting, pulled from Reddit, GitHub discussions, and Hacker News since the April 4 cutoff:

• Light personal use (20 messages/day, Haiku): $3-8/month. Basically fine.
• Moderate personal use (100 messages/day, Sonnet): $15-45/month. Noticeable but livable.
• Power user (500 messages/day, Opus): $60-200/month. This is where it hurts.
• Always-on agent with default heartbeat (Opus): $250-400/month. The heartbeat alone accounts for most of this.
• Misconfigured or unmonitored: $1,000-3,600/month. Not hypothetical - these are reported numbers.

For context, those same power users were paying $20/month on Claude Pro or $200/month on Claude Max 20x. Some are looking at a 10x to 50x increase.

Five ways to cut your bill right now

If you're staying on self-hosted OpenClaw with Claude API, here's what actually moves the needle:

1. Fix your heartbeat interval

The single biggest cost lever. Set your heartbeat interval just under Anthropic's cache TTL (currently 1 hour). A 55-minute heartbeat reuses the cached context instead of re-processing it from scratch. One user cut their bill by 90% with this and a few other tweaks. Or disable heartbeat entirely when you're sleeping - openclaw config set heartbeat.enabled false.

2. Drop to Haiku for routine tasks

One Redditor moved 85% of their workflows to Haiku and rerouted heartbeat checks to a free model via OpenRouter. Their bill dropped from $140/month to $12. Opus is 5x the price of Haiku on input and 5x on output. If your agent is checking email and summarizing Slack, it doesn't need Opus for that.

3. Purge your session context

Long sessions are expensive sessions. If you're in an ongoing conversation that's been running for hours, the context window is massive and every API call re-sends all of it. Clear your session regularly. Some people set up automated session resets every few hours.

4. Use OpenRouter for model routing

OpenRouter lets you route different tasks to different models based on complexity. Heartbeat and monitoring go to Haiku or a free model. Actual conversations go to Sonnet. You only hit Opus when you genuinely need it. This is more setup but it's the most effective cost control short of managed hosting.

5. Set a hard budget cap

Anthropic's API console lets you set spending limits. Do this before anything else. A runaway agent with no budget cap is how you get a $3,600 surprise. Set it to whatever you're comfortable losing in a month, then work backward from there.

The credit and the catch

Anthropic's offering a one-time credit equal to your monthly subscription price - $20 for Pro, $100 for Max 5x, $200 for Max 20x. Redeemable by April 17. They're also offering up to 30% off pre-purchased token bundles.

That $20 credit covers roughly 4 million Opus input tokens. At default heartbeat rates, that's about two days.

Some users on Reddit are reporting the credit redemption doesn't work. So check yours.

The switch-to-OpenAI math

Many people are moving to OpenAI, which officially supports OpenClaw through Codex OAuth. GPT-5.4 isn't Opus - multiple users admit the output quality is "not as human" - but OpenAI's subscription still covers it.

That's the short-term fix. The medium-term risk is OpenAI doing the same thing. Steinberger joining OpenAI gave them goodwill with the OpenClaw community. But Anthropic also supported OpenClaw at first. The pattern is: support the ecosystem until the economics don't work, then cut it off on a Friday afternoon.

If your workflows depend on a specific model provider's subscription terms, you're one policy email away from rebuilding everything. That's happened three times in four months now.

What a managed service actually solves here

The cost problem isn't just "tokens are expensive." It's that the cost is unpredictable, the defaults are wasteful, and provider policies change without warning.

A managed OpenClaw service handles all of this. Heartbeat intervals tuned for cost efficiency. Context management that doesn't let sessions balloon. Model routing that sends cheap tasks to cheap models. Provider relationships that don't depend on a single subscription's terms of service. And a fixed monthly price you can actually budget for.

TryOpenClaw.ai runs your OpenClaw agent for $39/month. No API keys to manage, no heartbeat tuning, no Friday afternoon policy emails. Just your agent, running, on channels you already use.