Claude Code Blank Thinking Blocks: Why You Still Pay (2026)
Claude Code users are seeing blank thinking blocks, yet Anthropic still bills every reasoning token behind them. Here is what happened, why your AI bill is affected, and the exact settings you can change today to stay in control.
📰 What Happened: Claude Code Shows Nothing but Bills Everything
On August 14, 2026, The Register reported that Claude Code, Anthropic's popular AI coding assistant, has been returning empty or truncated thinking blocks while still charging users for the reasoning tokens generated behind the scenes. Thinking blocks are the visible summaries of the model's step-by-step reasoning, the part where you watch the AI work through a problem before it answers.
Developers began filing reports around mid-July 2026. One GitHub report timestamped the problem at July 16, 2026, and noted that even when users explicitly requested summarized thinking, the tool returned blank content. Another report described summaries being silently cut off mid-stream. In the reporter's words, the thinking is generated and billed in full, while a portion of the summary stream is silently dropped.
The affected models named in the article are Claude Opus 4.8 and Claude Sonnet 5, two of Anthropic's most capable current models. At least four GitHub issues track the related problems, including #78200 for empty thinking blocks, #84780 for truncated summaries, #85322 for stream termination, and #86210 for the VS Code extension.
What is a thinking block, in plain English?
When you give Claude a hard task, it can reason privately before answering, similar to a person sketching notes before writing a final email. Extended thinking makes that scratchpad partially visible as a summary. The bug means the scratchpad still runs, and still costs money, but the summary you were supposed to see comes back empty or cut off.
💸 Why This Matters for Solopreneurs and Non-Developers
If you pay for AI by usage, this story is really about billing transparency. Anthropic's own documentation states that you are charged for all thinking tokens generated, even when collapsed or redacted. That policy is not new, and it is not secret. The problem is the combination: a bug hides the output while the meter keeps running, so you pay for work you cannot inspect.
For a solopreneur running Claude Code on API billing, thinking tokens are billed as output tokens, which are the expensive kind. On long, complex tasks, reasoning can quietly become a meaningful share of your bill. When the thinking display works, you can at least see that the model did substantial work. When it comes back blank, you have no easy way to judge whether the spend was justified.
There is a broader lesson here for anyone building a business on AI tools. Modern reasoning models from every major vendor, including Anthropic's Claude line, OpenAI's o-series, and Google's Gemini reasoning modes, all separate what the model thinks from what you see. As these tools become daily infrastructure for small businesses, understanding what you are billed for matters as much as understanding what the tool can do.
This is a display bug, not a new charge
To be fair to Anthropic, the reports describe a display and streaming failure, not a pricing change. You were always paying for thinking tokens. The bug removed your visibility into them, which is why users noticed and got frustrated.
🧮 How Thinking Token Billing Actually Works
Here is the mental model that makes this whole story click. Every response from a reasoning-enabled Claude model has two parts: the hidden or summarized reasoning, and the final answer. Both count as output tokens, and both count toward your maximum token limit for the request.
The table below compares what you see against what you pay for, based on Anthropic's documented behavior and the bug reports covered by The Register. Note that subscription users of the Claude Code plans experience this differently from API users, since API users pay per token while subscribers hit usage limits faster.
| Scenario | What You See | What You Pay For |
|---|---|---|
| Normal extended thinking | A summarized thinking block plus the answer | All thinking tokens plus answer tokens |
| The current bug | A blank or truncated thinking block plus the answer | All thinking tokens plus answer tokens, unchanged |
| Thinking disabled | Only the answer | Only answer tokens, no thinking cost |
| Stream cut mid-thinking | A partial or failed response | Tokens generated before the cutoff |
✅ What You Can Do Today: 5 Practical Steps
You do not need to be a developer to act on this. The Register's coverage points to two levers Anthropic itself recommends: lower the thinking budget or disable thinking entirely for tasks that do not need deep reasoning.
The key judgment call is matching thinking to task difficulty. Drafting a routine email or renaming files does not need extended reasoning. Debugging a gnarly spreadsheet formula or planning a multi-step automation probably does. Paying for thinking on hard problems is often worth it, because reasoning models genuinely perform better with room to think. Paying for it on trivial tasks is waste.
If you use Claude Code through a subscription
You are not billed per token, but heavy thinking still consumes your usage allowance faster. The same advice applies: reserve extended thinking for hard problems so your session limits last longer.
- ✔Check your usage dashboard at console.anthropic.com if you pay per token, and note how much of your spend is output tokens
- ✔Lower the thinking budget setting for routine tasks, since smaller budgets cap how many reasoning tokens can be generated
- ✔Disable extended thinking entirely for simple, repetitive work
- ✔Keep Claude Code updated, because Anthropic ships fixes nearly every day
- ✔If you see blank thinking blocks, check the open GitHub issues (#78200, #84780, #85322) before assuming your setup is broken
🏢 What Anthropic Says and What Comes Next
Anthropic did not dismiss the reports. A spokesperson told The Register that Claude Code ships updates nearly every day, and that community reports like these GitHub issues are a big part of how the company catches problems quickly. That is a reasonable description of how the tool actually operates, since Claude Code has an unusually fast release cadence.
On the separate stream termination problem, where the API cuts off long-running thinking sessions, the company said the issue relates to tuning network behavior to balance perceived latency against the risk of requests getting stuck. Translation: Anthropic tightened timeouts so requests feel faster and do not hang forever, and that tuning is sometimes cutting off legitimate long reasoning sessions.
For readers, the practical takeaway is that these are active, acknowledged bugs with public trackers, not a silent policy change. Watch the linked GitHub issues for resolution, and expect fixes to arrive through regular Claude Code updates rather than a dramatic announcement.
🔭 The Bigger Picture: Invisible AI Costs Are the New Normal
This story is a preview of a debate that will define AI tools in 2026 and beyond. As reasoning models become standard, a growing share of what you pay for happens off-screen. Anthropic summarizes thinking rather than showing it raw, OpenAI's reasoning models hide most of their chain of thought, and every vendor bills for the hidden portion.
Vendors have real reasons for this, including protecting their techniques and preventing misuse of raw reasoning traces. But it creates an accountability gap: when you cannot see the work, you cannot audit the bill. Incidents like this one, where a bug erases even the summary, make that gap visible and push users to demand better cost reporting.
If you run a business on AI tools, treat this as a prompt to build one small habit: review your AI spend weekly, know which tasks trigger expensive reasoning, and prefer tools that show you where the tokens go. The businesses that manage AI costs deliberately will have a durable edge over the ones that treat the bill as a surprise.
❓ Frequently Asked Questions
Is Anthropic overcharging Claude Code users?
No evidence suggests that. Billing for thinking tokens is documented policy, and the reported problem is a display and streaming bug that hides the thinking summary you were supposed to see. You are paying what you always paid, but temporarily losing visibility into part of it.
How do I stop paying for thinking tokens in Claude Code?
You can lower the thinking budget setting or disable extended thinking entirely, which is Anthropic's own advice for reducing thinking-related costs. The tradeoff is that complex tasks may get worse results without room to reason, so disable it selectively rather than globally.
Which Claude models are affected by the blank thinking blocks bug?
The Register's report names Claude Opus 4.8 and Claude Sonnet 5, with developer reports starting around mid-July 2026. Related problems are tracked in public GitHub issues including #78200 for empty blocks, #84780 for truncated summaries, and #85322 for stream termination.
Do ChatGPT and Gemini also charge for hidden reasoning?
Yes, this is industry-wide. OpenAI's reasoning models bill hidden reasoning as output tokens, and Google's Gemini reasoning modes charge for thinking as well. Any usage-based reasoning model charges for computation you mostly cannot see, so the cost-management habits in this post apply across vendors.
🏁 Final Thoughts
The short version: Claude Code has a bug that returns blank or truncated thinking blocks on Claude Opus 4.8 and Claude Sonnet 5, while the reasoning behind them is still billed as output tokens, exactly as Anthropic's documentation says it will be. It is a transparency failure, not a pricing scandal, and Anthropic has acknowledged the GitHub reports and is shipping fixes through its near-daily updates. Your move today is simple: match thinking budgets to task difficulty, disable extended thinking for routine work, and check your usage dashboard so hidden reasoning never surprises you on a bill. If you found this explainer useful, subscribe to Agents at Work for plain-English breakdowns of AI news that actually affects your workflow, and drop a comment telling us whether hidden reasoning costs have shown up in your own AI bills.
Last updated: August 14, 2026 · Keyword: Claude Code blank thinking blocks · Agents at Work
Comments
Post a Comment