Claude Code stops prompt suggestions spending your credits, and stable has neither fix
Claude Code was spending usage credits on its own prompt suggestions, the follow-up prompts it offers after a reply, whenever the model in use is one a plan can pay for only with credits. Anthropic's changelog says version 2.1.302 fixed it and that suggestions "are now skipped after a reply billed to credits", and the npm registry puts that release at 16:48 UTC on October 16.
Claude Code in a terminal, from Anthropic's own product page, where prompt suggestions appear once a reply is finished. Image: Anthropic
It is the second fix of this kind in two days. Version 2.1.302 went up 26 hours and 14 minutes after 2.1.301, which stopped a model switch spending credits on a summary by the model you were leaving.
The trigger and the mechanism are different, and in both cases Claude Code bought something on a user's behalf without putting the question first.
Stable is on 2.1.292 and has neither fix
The npm tags, read on October 18, put stable at 2.1.292, published on October 6 at 17:10 UTC, with latest and next both on 2.1.302. That is ten version numbers and eight published releases, because 2.1.297 and 2.1.298 were never published at all.
The eight that did ship ran from 2.1.293 on October 7 to 2.1.302 on October 16, and none of them has been promoted to stable.
Anyone pinned to that tag keeps the behaviour. Suggestions after a credit-billed reply still get generated and still get charged on a machine running 2.1.292, and the gap is a familiar one: stable sat four builds behind a fix to the guard on destructive shell commands earlier this month.
Anthropic's review view for a submitted Claude Code plugin. The same release cuts the file system calls needed to load plugin commands and agents. Image: Anthropic
Usage credits are spent past a plan's limit
Usage credits are the balance that lets a subscriber keep working past a plan's usage limit, according to Anthropic's documentation, managed with /usage-credits and switched on, capped and tracked from the usage settings on claude.ai, or by whoever holds billing access on a Team or Enterprise plan.
That command is not available when Claude Code is signed in with an API key, where usage is billed per token anyway. So the charge the fix removes landed on money a subscriber had agreed to spend past the plan's allowance, rather than inside it.
Claude suggesting threads to start. Prompt suggestions in Claude Code do the same job, and until 2.1.302 they could be charged to credits. Image: Anthropic
Anthropic publishes no numbers for either case
The changelog does not say how much was charged, how many accounts saw it, when the behaviour started or whether anything is refunded, and it says "usage credits" rather than dollars. Anthropic's status page records no incident on October 15, 16 or 17.
The suggestions themselves were real output, generated and delivered. What 2.1.302 changes is that they are skipped after a reply the credits paid for, so the charge cannot arrive unasked.
The same release moves a permission default
Anthropic says claude -p and Agent SDK sessions started with no permission mode now begin in default mode on third-party providers, with telemetry off or on a fresh install, while --permission-mode auto keeps auto.
Plugins get two changes in the same release. Anthropic says plugin command and agent files now take far fewer file system calls to load, which it lists as a faster start for sessions with many plugins enabled, most of all on Windows, and it fixed plugin marketplaces re-cloning on every refresh when git defaults to the reftable ref format.
The rest of 2.1.302 is long, with fixes for MCP servers that negotiate the 2026-07-28 protocol without the fields Claude Code expects, for file writes being refused on exFAT and FAT32 drives on Windows, and for a scheduled cloud run that failed partway being started again from the beginning an hour or more later.
None of it reaches the stable tag, which has not moved since October 6.
