How I Cut My OpenClaw Costs by 94% After Anthropic Killed Claude Subscriptions

The Problem: Anthropic Locks Out OpenClaw Users

On April 4, 2026, Anthropic announced that Claude Pro and Max subscribers can no longer use their subscription with third-party AI agent frameworks like OpenClaw. If you want to keep using OpenClaw with Claude, you now need to pay separately through their “extra usage” pay-as-you-go billing — and users are reporting 10x to 50x cost increases compared to what they were paying before.

Boris Cherny, Head of Claude Code at Anthropic, justified the move by saying “Anthropic’s subscriptions weren’t built for the usage patterns of these third-party tools.” OpenClaw creator Peter Steinberger didn’t mince words: “First they copy some popular features into their closed harness, then they lock out open source.”

With over 135,000 OpenClaw instances running at the time of the announcement, this hit a lot of people — myself included.

I rely on OpenClaw for deep research tasks through WhatsApp. I wasn’t about to start paying $1-5 per research session on Claude’s API directly. So I set out to find a cheaper alternative that didn’t sacrifice too much quality.

Step 1: Understanding the Damage

I was already using OpenRouter as my provider (since the Claude subscription route was cut off). I downloaded my OpenRouter activity report and the numbers were brutal:

One 10-minute research session on April 6:

The culprit was obvious. My OpenClaw was set to openrouter/auto — OpenRouter’s auto-router that picks models based on prompt complexity. For my deep research queries, it kept choosing Claude Opus 4.6 at $5/M input tokens, $25/M output tokens.

Worse: 7 agentic loop calls in 10 minutes, each sending ~32,000 prompt tokens to get back an average of just 214 completion tokens. Five of those calls were just tool-call dispatches — the agent thinking about which tool to use next. At $0.16 per round-trip, that adds up fast.

And to top it off: zero prompt caching. Every single Opus call had tokens_cached = 0. The GPT-5 Nano calls were caching tokens automatically, but Anthropic’s Claude requires explicit cache_control markers that OpenRouter doesn’t add for you. Without caching, I was paying full price on ~30K tokens of identical context repeated 7 times.

Step 2: The Search for Alternatives

I didn’t want to sign up for 10 different AI provider accounts, manage separate API keys, and test each one individually. That’s the whole point of OpenRouter — one API key, access to hundreds of models.

So I dug into what’s actually available on OpenRouter and cross-referenced with the Artificial Analysis LLM Leaderboard for quality benchmarks.

The Full Comparison

Here’s what the landscape looks like — every serious model, side by side, sorted by value:

Key insight: Claude Opus 4.6 scores 53 but costs $10/M blended. MiniMax-M2.7 scores 50 (just 6% lower) at $0.53/M blended — that’s 19x cheaper for 94% of the quality.

Models I Seriously Considered

MiniMax-M2.7 — The clear winner. Score of 50 matches Claude Opus 4.5. Massive 205K context window. $0.30/M input. For deep research where you’re feeding in long documents and asking complex questions, this is the sweet spot.

Qwen3.5 397B — The best open-source model. Score 45, 262K context (largest among cheap models), and vision support for analyzing charts and images. At $0.39/M input, it’s excellent for document-heavy research.

DeepSeek V3.2 — The absolute cheapest model worth using. Score 42 at just $0.20/M input. Great for analytical and mathematical reasoning. If you’re on a tight budget, this punches way above its weight.

MiMo-V2-Flash — Ultra-budget option from Xiaomi. Score 41 at $0.09/M input. 262K context. For quick lookups and simple tasks, you can’t beat it.

Qwen3 235B :free — Literally free on OpenRouter. Score 25 isn’t amazing, but for casual research and experimentation, zero dollars is zero dollars. Rate limited to ~20 requests/minute.

Step 3: What I Actually Set Up

Here’s my final OpenClaw configuration:

Primary model: openrouter/minimax/minimax-m2.7

• Best bang for buck at score 50
• Handles deep research queries well
• 205K context for long documents

Fallbacks: DeepSeek V3.2 → Qwen3.5 397B (auto-switches if MiniMax is down)

All models available via /model command in chat:

MiniMax M2.7      — daily driver, best value (score 50, $0.30/M)
Qwen 3.5 397B    — long docs, vision support (score 45, $0.39/M, 262K ctx)
DeepSeek V3.2    — cheapest smart model (score 42, $0.20/M)
MiniMax M2.5     — budget option (score 42, $0.12/M)
MiMo Flash       — ultra cheap quick tasks (score 41, $0.09/M)
Gemini 2.5 Flash — massive 1M context window ($0.30/M)
Qwen 235B FREE   — zero cost, rate limited
Claude Opus      — nuclear option, when nothing else works

The relevant openclaw.json config:

"model": {
  "primary": "openrouter/minimax/minimax-m2.7",
  "fallbacks": [
    "openrouter/deepseek/deepseek-chat-v3-0324",
    "openrouter/qwen/qwen3.5-397b-a17b"
  ]
}

The Result

For deep research, the 6% quality difference is barely noticeable. The model still reasons well, follows complex instructions, handles tool use, and synthesizes information from long documents. The only time I’d switch back to Opus is for genuinely novel, frontier-difficulty reasoning tasks — and even then, I’d try Qwen3.5 397B first.

Recommendations

If you’re an OpenClaw user dealing with the same cost shock:

1. Stop using openrouter/auto — it will keep routing your complex queries to Opus and draining your wallet.
2. Set MiniMax-M2.7 as your primary — 94% of Opus quality at 6% of the cost.
3. Set a budget cap on OpenRouter — go to your OpenRouter dashboard, set a daily or monthly limit on your API key. Even $5/day is plenty with these models.
4. Use /model to switch contextually — doing a quick lookup? Use MiMo Flash at $0.09/M. Analyzing a 100-page PDF? Use Gemini 2.5 Flash with its 1M context. Need the absolute best? Opus is still there.
5. If you’re using Claude via OpenRouter, enable prompt caching — add cache_control markers to your system prompts. Anthropic cache reads cost 0.1x input pricing. This alone can save 80%+ on agentic loops where the context barely changes between calls.

The AI model landscape has gotten incredibly competitive. You no longer need to pay frontier model prices for frontier-quality results. The gap between a $0.30/M model and a $5.00/M model has shrunk to single digits on quality benchmarks — and for most real-world research tasks, that gap is invisible.

Written with the help of Claude Code, which ironically still runs on Claude Opus. The cobbler’s children go barefoot.