A developer-driven investigation published as a Gist on September 5, 2026, alleges that Anthropic’s Claude models are billing users for 'extended thinking' tokens without providing the corresponding reasoning traces in the API response. The post, titled 'You're paying for Claude's thinking and you're not getting it,' argues that the value proposition of the premium 'thinking' mode is being undermined by opaque implementation.

The 'Black Box' Billing Complaint

The core of the argument rests on the discrepancy between cost and transparency. Extended thinking was marketed as a way to see the model's chain-of-thought, allowing developers to debug complex logic and trust high-stakes outputs. However, the Gist suggests that while the token count reflects the additional computational effort of 'thinking,' the actual intermediate steps are either truncated, omitted, or inaccessible in standard API calls. This effectively turns a 'glass box' feature back into a black box, but with a higher price tag.

Technical Implications for Developers

For builders relying on Claude for agentic workflows, the visibility of reasoning is not just a nice-to-have—it is a critical debugging tool. If the API charges for tokens generated during the thinking phase but does not return them, developers lose the ability to audit why a model chose a specific path. This creates a scenario where costs scale with 'thinking' depth, but observability remains static. The community is left to infer whether the model is actually performing deeper analysis or if the token billing is simply inflated for the same output quality.

Source Limitations and Verification

It is important to note that the source material provided consists primarily of binary-compressed text from the original Gist, which makes extracting specific code snippets or exact API response examples difficult without decompression. However, the headline and the Hacker News discussion context clearly point to a widespread sentiment among power users that the 'thinking' feature is not delivering on its transparency promise. The post has currently gathered 4 points on Hacker News with zero comments, suggesting it may be a fresh or niche complaint, but the technical assertion aligns with ongoing debates about LLM pricing models.

Key Takeaways

  • Developers are reporting that Claude's 'extended thinking' mode charges for reasoning tokens but fails to expose the reasoning steps in the API response.
  • The lack of visible chain-of-thought output undermines the primary utility of the feature for debugging and auditing agentic workflows.
  • This issue highlights a growing tension between premium pricing for advanced LLM features and the need for developer observability.

The Bottom Line

If you are paying a premium for 'thinking,' you should see the thoughts. Anthropic needs to clarify whether this is a bug, a documentation gap, or a deliberate obfuscation strategy—because right now, it feels like we are being billed for invisible labor.