Claude Code could spend your credits on a summary by the model you were leaving, until 2.1.301

Switching Claude Code to a model with a smaller context window could spend your usage credits on a summary written by the model you were leaving, without asking first. Anthropic’s changelog says version 2.1.301 fixes that case, and the npm registry puts the release at 14:34 UTC on October 15.

The summary itself is deliberate. A conversation too long for the smaller model is summarised by the previous model first, which Anthropic lists as an improvement in the same release, and what it has fixed is the case where those credits went without a prompt.

Anthropic attaches no figures. The changelog does not say how much was charged, how often it happened, which models were involved or how long the behaviour had been there, and it says “usage credits” rather than dollars.

The mechanics are easy enough to picture. A long conversation will not fit into the smaller model’s window, so something has to compress it, and the only model holding the full conversation at that moment is the one you are stepping away from.

Charging and credential faults have run through Claude Code’s October releases. Version 2.1.299 stopped sending a third-party gateway’s API key to Anthropic on October 14, and in September Claude Code stopped charging API users for auto mode’s safety checks.

A gateway could trigger repeated re-compaction

A separate fix in 2.1.301 closes a different fault, on a different trigger. Anthropic says it fixed “conversations being compacted over and over, on a model you didn’t pick, behind an LLM gateway whose responses name a different model than the one requested”.

That condition is specific. A gateway sits between Claude Code and a model provider so an organisation can log, cap and restrict what its developers send, and the loop needed the gateway’s replies to name a model other than the one asked for, which puts it on teams rather than on anyone calling the API directly.

Model switching picked up a third repair in the same release. Anthropic says a model name containing application-inference-profile was running a developer’s own awsAuthRefresh command and delaying model-switch hooks even when the provider was not Bedrock.

A fable alias now retries on an older model

Another fix in the release documents something about model availability. Anthropic says the fable alias, and a best alias alongside it, was erroring on the Anthropic API when the newest Fable model is not available to an organisation, and that Claude Code now retries once on the previous Fable model.

So the newest Fable model is not something every organisation can reach, and until this release an alias pointing at it failed rather than falling back.

One more entry in 2.1.301 is about waiting rather than spending. Anthropic added a setting, CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS, so unattended retry mode ends a request when the usage limit it hits resets too far away, instead of sitting there until the reset arrives. The release also puts a “Get more usage” link in the usage-limit banner that Pro and Max users see in the VS Code and JetBrains panels.

Three builds in two days, and stable has none of them

Anthropic published 2.1.299 at 00:23 UTC on October 14, 2.1.300 at 14:27 UTC the same day and 2.1.301 the following afternoon, according to the npm registry. Versions 2.1.297 and 2.1.298 were never published at all.

None of it reaches a machine pinned to the stable channel. That tag still points at 2.1.292, published on October 6, seven published releases back.

Note: Links to online stores in articles can be so-called affiliate links. If you buy through this link, MIXED receives a commission from the provider. For you the price does not change.