Claude Code blank thinking blocks: meter your own tokens

Since mid-July, Claude Code and the Anthropic API have been returning thinking blocks with an empty thinking field — a signature present, the summary gone — on Claude Opus 4.8 and Sonnet 5, even when callers explicitly ask for summarized display. The reasoning still runs. The reasoning still bills. You just can't read it. That combination is worth more attention than the bug itself.

What actually happened

The Register reported on August 14 that developers began seeing empty thinking summaries around July 16, with reports piling into the tracker. GitHub issue anthropics/claude-code#63358 describes Opus 4.8 returning thinking blocks whose content field is blank, so the UI renders nothing; issue #56356 logs the same behavior on Opus 4.7 with --thinking adaptive --thinking-display summarized set.

An Anthropic spokesperson told The Register the reports are being investigated, noting that Claude Code "ships updates nearly every day" and attributing some of the behavior to tuning network handling around long-running requests.

The billing side is not in dispute and does not require a bug report. Anthropic's own thinking documentation tells you to watch usage.output_tokens_details.thinking_tokens, which it describes as reporting how many of the billed output tokens were internal reasoning. Thinking is output. It is metered whether or not anything renders on your screen.

Why blank thinking blocks matter for your business

Treat this as a preview, not an incident. The trend across every frontier vendor is toward reasoning that is summarized, encrypted, or hidden — for competitive and safety reasons that are entirely defensible. The consequence for anyone paying the bill is the same either way: the most expensive part of a request is becoming the least inspectable part.

If your only view of agent spend is a dashboard the vendor renders, you are auditing with their instrument. Read the usage object yourself. Every response carries the token breakdown; log thinking_tokens per request alongside a job identifier, and you get a cost-per-task series that survives any UI change, any display regression, and any pricing page rewrite. That is a few hours of work and it is the difference between "our AI bill went up" and "the invoice-parsing agent's reasoning tripled after Tuesday's prompt change."

Second, stop treating the thinking summary as your debugging surface. It was never a faithful trace of the model's reasoning — it was a rendered summary, which is exactly why it can vanish without anything else breaking. If you need to know why an agent did something, the durable answer is your own logs: inputs, tool calls, outputs, and the diff it produced. Build those and a blank thinking block is an annoyance instead of an outage.

Third, put the boring line in your vendor notes. Daily-shipping tools are a real advantage — we wrote about Claude Code shipping eight permission fixes in a day — but the same cadence means display behavior is not a contract. Pin what you can, instrument what you can't.

Sources: The Register, August 14, 2026, anthropics/claude-code issue #63358, Anthropic extended thinking docs.