claude-code v2.1.322

Claude Code 2.1.322 lets people who run Anthropic's coding agent through Google Cloud's Vertex AI turn on fast mode, and when Vertex will not serve a request at fast speed, Claude Code sends it again at standard speed instead of failing (changelog). Fast mode runs the same model with a serving configuration that, according to Anthropic's fast mode documentation, produces up to 2.5 times more output tokens per second at a higher price, without changing the model's answers or abilities. The gain is in how fast text streams out, not in how soon the first token arrives, so it shows most on long responses such as large edits. Until now, Claude Code's own fast mode page limited the feature to Anthropic's API and Claude subscriptions; as read on November 10 it still says fast mode is not available on Google Cloud's Agent Platform, the new name for Vertex AI, so the documentation has not caught up with this release.

The three regions in the release note match how Claude Code addresses Vertex. Google Cloud offers a global endpoint that sends requests to any region with capacity, us and eu multi-region endpoints that keep requests inside the United States or the European Union, and single-region endpoints such as us-east5 (Anthropic's Vertex documentation). Claude Code picks one through its CLOUD_ML_REGION setting. The same documentation says single-region endpoints serve only Claude Sonnet 4.6 and older models, none of which support fast mode, which fits fast mode being offered on global, us and eu only. Multi-region endpoints cost 10 percent more than global ones.

The fallback keeps a session working, but it is not free. A request at fast speed can be refused because the model or account is not enabled for it, or because fast capacity or rate limits are exhausted, which the API reports as a 429 or 529 error. Fast and standard speed keep separate prompt caches, so a request resent at standard speed misses the cache, and part or all of the conversation so far is read again at the uncached rate. Billing also differs from Anthropic's own API: on Vertex, Google Cloud sets fast mode prices and manages quotas, so the $8 and $40 per million tokens that Anthropic lists for Claude Opus 5.5 fast mode are not necessarily what a Vertex customer pays.

One gap is worth checking before relying on the feature. Anthropic's availability table lists Google's Vertex offering for fast mode on Claude Fable 5.2 and Claude Mythos 5.2, while its Claude Opus 5.5 row names only the Claude API, the Anthropic-managed Google Cloud offering and Microsoft Foundry. Opus 5.5 is Claude Code's default model on Vertex, and the release note does not say which models it covers, so a Vertex user on Opus 5.5 may find that fast mode simply falls back to standard speed. When Anthropic made Opus 5.5 fast mode generally available on its API in October, that announcement said Google Cloud support was coming soon; this release brings fast mode to Claude Code users on Vertex, though which models it serves there is the open question above. The previous version, 2.1.321, had already changed fast mode to try standard speed on the same model first when fast serving is rate-limited, overloaded or failing.