Some developers are using the Claude Code harness to access cheaper non-Anthropic models, such as OpenAI's GPT-5.6 Sol, via proxies and services like OpenRouter
Boris Cherney, the head of Anthropic's coding tool Claude Code, last month found himself in the middle of an online spat that highlighted …
Context & Ripple Effects
Claude Code moved from general availability in 2025 to a product credited with helping establish Anthropic in AI coding tools. Anthropic had already added anti-spoofing safeguards against third-party apps seeking more favorable access to its own models.
The proxy use described here reverses that access problem: developers keep Claude Code as the working interface while directing requests to other model providers. That separates the value of the coding harness from Anthropic's token sales.
First-order effects
- Developers using proxies or OpenRouter can retain Claude Code workflows while choosing cheaper non-Anthropic models, including GPT-5.6 Sol.
- Anthropic forgoes token revenue on requests routed through its harness to external models rather than Claude.
Second-order effects
- OpenRouter and the providers behind alternative models gain demand generated by Claude Code users without needing to displace the harness itself.
- Anthropic faces pressure to make its pricing, access rules, or product integration competitive with the cost of using the same workflow through a router.
Third-order effects
- If this pattern persists, AI coding tools may divide into a durable control layer and a contestable inference layer, with model providers competing for traffic beneath familiar developer interfaces.
- Usage controls become more consequential: providers will need to distinguish legitimate multi-model routing from attempts to obtain restricted or discounted model access.
The trend: AI access arbitrage is turning coding agents into model-agnostic control planes, shifting competition from the interface toward inference price and routing terms.