AINEWS
Anthropic's paid coding assistant stopped showing its internal reasoning, while the tokens behind that reasoning remain on the user's invoice.
Anthropic's Claude Code used to pay its subscribers a small but real courtesy: it showed them the model's internal reasoning before each answer, in a visible "thinking" block that scrolled out above the final code. That block now comes back empty for many users, a regression Anthropic has not publicly resolved. The tokens behind the silence are still on the bill.
Claude Code is Anthropic's paid AI coding assistant, a command-line tool aimed at software developers that ships answers alongside the reasoning trace used to produce them. The trace was the part a paying user could audit. On the 2.1.69 build, the block returns with only an encrypted signature and thinking_len=0, a placeholder where the reasoning used to be, according to a user report filed in the public anthropics/claude-code issue tracker (issue #31326). A second, parallel report on Opus 4.8 explicitly links the new behavior back to an earlier Opus 4.7 regression (issue #63358, issue #49268), suggesting the team has been chasing this surface across more than one model release.
The visible block was never the thinking itself. Anthropic's platform documentation on extended thinking describes the feature as a server-side process the API returns as a separate content block alongside the model's answer; the CLI was simply rendering that block for the user. When the block returns empty, the model has not stopped reasoning. A plan caught in the visible trace is a plan that never costs the user a finished answer; the user has lost the cheap signal that the model had picked a sane path before it started writing code.
A user watching a scroll of model reasoning could do two things at once: read the plan, and check whether the plan was sane. With the block gone, the user still gets the code, still pays the tokens, and has lost the only lightweight audit the product ever offered. The cost surface (the meter) and the inspection surface (the visible block) have been decoupled by a rendering change that did not touch the price. The Register re-reported the regression this week and framed it as a thinking-still-billed story. The framing survives a careful read, but any specific dollar figures are not in the public tracker, and they should not be invented here. What the public record does carry is the regression itself, the build it shipped in, and Anthropic's own docs describing what the visible block is supposed to surface.
A paying user has more than a complaint to file. The token-usage dashboard still records every reasoning token Claude Code spent, so the invoice can be cross-checked against the work the user actually saw. Independent explainers describe a workaround pattern of turning thinking back on, which suggests the feature is configurable rather than purely broken (hypogray.com, wmedia.es). Anthropic's own support channel is a natural place to ask for visible reasoning on a specific run. If a fix does not land, the lever that matters is pricing: a paid tool that no longer lets a customer inspect what it billed for is charging for a window that has been painted over.
This is, on Anthropic's behalf, a temporary rendering bug, and the model is still reasoning on the server. If the company publishes a restore timeline and ships it, the story softens to "shipped a regression, will fix." That answer covers the surface; it does not cover the cost-stays-the-same observation, or the inspection-surface / cost-surface split the regression made visible. A dated fix in the public issue tracker would turn the trust question into a closed bug; the watch item is whether that fix arrives.