Anthropic Cut Off OpenClaw from Claude. How to Switch to Free Models

What Happened with Anthropic#

Here’s how it used to work: you paid for a Claude subscription ($20 – $200/month). OpenClaw used the same models via OAuth. The subscription covered everything – both the chat and the server.

Now the subscription works only in Anthropic’s own products. Claude.ai, Claude Code, Claude Desktop. Period. Third-party applications have been moved to “Extra Usage” – a separate per-request billing system.

Here’s what you see in chat when the bot goes silent:

WARNING

“LLM request rejected: Third-party apps now draw from your extra usage, not your plan limits.”

Why did they do this? Arbitrage. A $200 subscription was covering $1,000–$5,000 worth of compute. Anthropic was subsidizing AI servers thinking everyone was just chatting. OpenClaw’s agent loops consume 5–25x more tokens than a regular conversation. The math didn’t work.

HackerNews was split down the middle. What caught my attention was the timing.

Qwen 3.6 Plus Launched Two Days Before the Ban#

On April 2, Alibaba released Qwen 3.6 Plus. A free model with a million-token context window. On April 4, Anthropic shut down OAuth. Coincidence? Maybe. But the market reacted instantly.

I’m looking at OpenRouter stats – the largest AI model routing platform – and here’s what I see:

Programming: Qwen took 84%.

Not a typo. 83.7% of all programming requests go through Qwen. Claude Opus and Sonnet are somewhere at the bottom. This isn’t an opinion poll. Thousands of developers voted with their time and money.

Technology – 67%. Analytics – 55%. Russian language – also first place.

Across all OpenRouter categories – from code to Russian-language text processing – Qwen is in first place. Not by a small margin. By a landslide.

The key takeaway for me – people aren’t “trying” Qwen. They’ve moved. 84% is a migration, not a test.

Compensation Credits – Check Your Account#

Anthropic automatically credited compensation to affected accounts. The amount depends on your plan:

Credits are applied automatically – no action needed to claim them. You can use them for Extra Usage across Claude.ai, Claude Code, Claude Desktop, or third-party apps.

WARNING

Credits expire April 17, 2026. Check your balance at claude.ai/settings/usage and use them before they’re gone.

Can you cancel your Claude subscription? Yes. If you’re switching to free models, the subscription is no longer needed. Credits will stay in your account regardless.

Recovery in 10 Minutes#

You already have Claude Code – we installed it when setting up your server. It still works through your subscription. Let’s use it to fix OpenClaw.

1
claude --dangerously-skip-permissions

1
Find my OpenClaw server and do the following:
2

3
1. Add OpenRouter provider:
4
   - Get a free API key at openrouter.ai
5
   - Add it to openclaw.json as "openrouter" provider
6

7
2. Add NVIDIA NIM provider:
8
   - Get a free API key at build.nvidia.com
9
   - Add it to openclaw.json as "nvidia" provider
10

11
3. Switch the primary model:
12
   - Primary: openrouter/qwen/qwen3.6-plus:free
13
   - Fallback: nvidia/nemotron-3-super-120b-a12b:free
14

15
4. Restart the gateway.
16
5. Send a test message and verify it works.

Free Tier Limits#

There’s a catch. OpenRouter’s free tier is 20 requests per minute, 200 per day. Sometimes the bot may say the limit is exceeded.

TIP

Hit the limit? Wait a minute and try again. Sent a voice message? Forward it again – the bot will process it at the next available slot.

If limits are a constant problem – top up OpenRouter with $5–10. Limits are completely removed, and Qwen costs pennies. Ten dollars lasts months.

Right now Qwen is seeing explosive traffic growth – everyone is migrating. Alibaba is scaling servers. The situation will stabilize.

Three Configurations#

1
Orchestrator:  Qwen 3.6 Plus (OpenRouter)
2
Fallback:      Nemotron 120B (NVIDIA)

1
Orchestrator:  Claude Opus 4.6 (Anthropic API)
2
Sub-agents:    Qwen 3.6 Plus + Nemotron (free)

1
Orchestrator:  Claude Opus 4.6
2
Sub-agents:    Claude Sonnet 4.6 + Qwen 3.6 Plus

How Orchestration Works – Briefly#

Your AI server is a conductor. It receives a task, breaks it into parts, and delegates to performers.

Orchestrator – the lead agent. Thinks, plans, makes decisions. The CEO. Needs the smartest model.

Sub-agents – the performers. Search, write, analyze, compare. Free models work fine here.

Example. You say: “Prepare a competitor report.”

1. Orchestrator (Opus) breaks it into sub-tasks
2. Sub-agent (Qwen, $0) searches for competitors
3. Sub-agent (Qwen, $0) gathers prices and reviews
4. Sub-agent (Nemotron, $0) analyzes websites
5. Orchestrator synthesizes the final report

Opus quality. Cost – pennies. That’s hybrid orchestration.

You Have Claude Code. Use It.#

You pay for a Claude subscription. It includes Claude Code – a tool that runs in the terminal. It’s still available.

Claude Code can connect to servers, edit configs, run scripts, diagnose problems. If something breaks:

1
claude

1
Connect to my server, check OpenClaw status,
2
show the latest errors in logs and suggest a fix.

It’s your Swiss army knife. It works.

Where Things Are Headed#

Here’s how I see it: the Anthropic block is the best thing that could have happened.

Dependency on a single provider is a vulnerability. Anthropic just proved it. Today they change OAuth policy. Tomorrow they raise prices. The day after – something else.

The AI model market is no longer a monopoly. Qwen, Nemotron, DeepSeek, Gemma, Llama – dozens of models. Many free. Many excellent. And getting better every month.

We’re heading toward a future where you own your models. Running them locally, on your own hardware. No dependency on clouds or subscriptions. OpenClaw already supports this via Ollama. And if you want to start from scratch – here’s how to install OpenClaw.

But for now – free cloud models cover 90% of needs. And cost less than a $200 Claude Max subscription.

Switch over. Use your credits before April 17. Keep working.

FAQ#

Do free models actually work?

84% of developers on OpenRouter chose Qwen for code. 67% for technology tasks. People don’t mass-adopt something that doesn’t work. Qwen handles most tasks as well as Claude. Opus is better at non-trivial reasoning – but that’s what the $30 hybrid configuration is for.

Is my data safe?

OpenRouter and NVIDIA process requests on their servers. Security level is standard for any cloud API. For confidential data, I recommend Claude Opus via API or a local model through Ollama.

I don’t understand terminals.

Reach out – I’ll connect remotely and set everything up in 10 minutes.

Should I use my credits?

Absolutely. They expire April 17 – even if you’ve switched to free models, use them as a fallback while they last.

Need help with setup? Contact me – I’ll connect and configure everything.