Anthropic recently updated their terms of service to block third-party CLI tools from using Claude Code subscriptions for API access. This has sparked quite a debate in the developer community, and I wanted to share my thoughts on it.

The core issue is straightforward: Anthropic offers Claude Code subscriptions at rates significantly lower than their API pricing. A heavy user might consume over $1,000 worth of API tokens in a month while only paying $200.

Third-party tools like OpenCode and Crush discovered they could spoof the Claude Code CLI - sending matching headers and client signatures to make Anthropic’s servers believe requests came from the official tool - and access this discounted compute at subscription prices.

I had been using a third-party CLI harness too for unique workflows, so I get the frustration of losing access.

The way I see it, this falls on a fine line similar to security disclosure ethics. When security researchers discover vulnerabilities, the community has established clear norms: ethical disclosure is rewarded, while exploiting vulnerabilities for personal gain is condemned.

The parallel here is notable. Everyone in the developer community knew these were deeply discounted tokens intended for Anthropic’s official tooling. Using third-party workarounds to access this pricing wasn’t discovering a hidden feature - it was knowingly exploiting a pricing structure that wasn’t designed for that use case.

Although it’s understandable that clarity from Anthropic came rather late, and the confusion still remains because Agents SDK is quoted as allowed while docs claim it’s not.

That said, I don’t think Anthropic is being unreasonable here. Some argue they should have open-sourced Claude Code or offered API credits at a discount instead. While these might be preferable from a user perspective, they ignore the business reality.

Claude Code represents significant competitive advantage for Anthropic - it’s arguably what makes them relevant in the coding assistant space, beyond just the models themselves.

The situation highlights an interesting contrast with other players. GitHub’s Copilot team appears to be taking the opposite approach, actively working with the OpenCode team to support subscription authentication officially. This collaborative stance could set a different precedent for how AI companies handle third-party integrations.

For developers who relied on these workarounds, the path forward is clear: use the official Claude Code CLI, pay for API access separately, or explore alternative providers like Groq, Together AI, OpenCode Black, or Factory Max Plan.

It’s frustrating, but not unreasonable.

The broader lesson applies beyond just AI tooling - when companies offer favorable pricing for specific use cases, building dependencies on workarounds that exploit that pricing is inherently risky. The terms can change at any time.