Invisible Intellect, Visible Bills: Claude Code's Reasoning Riddle

In the fast-evolving world of AI, transparency is often touted as a cornerstone of trust and effective development. Large Language Models (LLMs) like Anthropic's Claude Code offer a glimpse into their "thinking" processes, providing invaluable insights for developers integrating these powerful tools. However, recent reports from the developer community suggest a concerning disconnect: Claude Code is returning blank thinking blocks, yet the underlying reasoning still costs you. This isn't just an interface glitch; it's a potential hit to budgets and a challenge to the very idea of transparent AI.

The Case of the Vanishing Thoughts

The core of the issue, as highlighted by multiple bug reports, centers on Claude Code's API for its Opus 4.8 and Sonnet 5 models. Developers, including software engineer Michael Hood, observed that as of July 16, 2026, the API began returning empty or severely truncated "thinking blocks." This occurred even when explicit requests for "summarized" thinking were made, and even when injected directly into the raw request body. Similar behavior has been noted in Claude Code for VS Code, indicating a broader problem.

Imagine building a complex application, relying on Claude Code to not just generate code, but to explain how it arrived at that solution. These thinking blocks are crucial for debugging, understanding the AI's logic, and ensuring the generated output aligns with expectations. When they come back blank, developers are left in the dark, paying for an invisible process.

Paying for the Unseen: Anthropic's Billing Policy

The real sting for developers lies in the billing. While the thinking blocks disappear, the meter keeps running. Anonymous reports claim that the thinking is generated and billed in full, with only the summary stream being silently dropped. This has led to frustration and a sense of being billed for undelivered value.

However, Anthropic's documentation offers a clear, if somewhat unyielding, explanation: "You are charged for all thinking tokens generated, even when collapsed or redacted." Furthermore, "Thinking has a cost: the tokens Claude spends reasoning are billed as output tokens, even when the thinking text isn't returned to you, and they count toward max_tokens alongside the response text." This legalese means that whether the summary is displayed, truncated, or completely absent due to a bug, the computational effort for that reasoning is still an expense.

Under Investigation: Balancing Latency and Transparency

Anthropic acknowledges the issue, stating it's under investigation. An Anthropic spokesperson told The Register that the problem might be an "artefact of tests that change how Anthropic displays summaries" and could be related to "tuning network behavior, specifically to terminate or retry long running requests." This suggests a complex balancing act between perceived latency and the risk of requests getting stuck.

While the company is "grateful to the developers who take the time to file them" and assures that "Claude Code ships updates nearly every day," the immediate financial impact on users remains. Until a definitive patch is issued, developers are essentially paying for a black box process.

Navigating the Costs: Advice for Developers

For enterprise developers building on Claude Code, vigilance is key. Until Anthropic stabilizes the summarized outputs, it's prudent to:

• Audit API Expenditure Logs: Regularly check your billing details against actual usable output.
• Consider Disabling Thinking: Anthropic advises customers to lower their budget setting or disable the thinking feature entirely to reduce costs associated with unseen reasoning tokens. This might temporarily sacrifice transparency for cost control.

This situation underscores the critical need for clear communication and robust delivery in AI services. When the "thinking" is a core feature, its absence – especially when still billed – erodes trust and complicates development. As AI models become more integral to our workflows, ensuring that what you pay for is what you get, and what you see, becomes paramount.