Source: ai-research/claude-code-docs-changelog-2026-09-25.md (changelog v2.1.278, September 19, 2026). No official W38 digest page was published at code.claude.com/docs/en/whats-new/2026-w38 as of 2026-09-25 (returns HTTP 404). This article is compiled from the changelog only.
W38 shipped a single notable release: v2.1.278 (September 19) which changes the auto mode classifier to run server-side by default for Claude API, Enterprise, Bedrock, Vertex, Foundry, and gateway deployments. Continues Week 37.
Key Takeaways
- Auto mode classifier moved server-side (v2.1.278) — for Claude API and Enterprise users, and on Bedrock, Vertex, Foundry, and gateway deployments, the auto mode classifier now runs server-side by default instead of locally. This eliminates the per-classifier-request charge that local classification was incurring. If server-side classification results in a billed fallback, Claude Code now warns you.
Auto mode serverrow in/status— after v2.1.278,/statusincludes a new row showing where the auto mode classifier is running (server-side vs. local). Useful for confirming which classification mode is active in your deployment environment.- Opt-out available — set
CLAUDE_CODE_AUTO_MODE_SERVER=0to revert to local classifier behavior if your deployment requires it. - v2.1.279 absent — the changelog skips from v2.1.278 to v2.1.280. This matches the established pattern of occasional skipped version numbers (v2.1.230, v2.1.262, etc.); no release with that version number appears in the changelog.
Auto mode server-side classifier
Before v2.1.278, the auto mode classifier that determines whether an action requires user approval ran locally inside the Claude Code process. This incurred a per-classification charge on accounts where the classifier was billed separately.
From v2.1.278 onward, the classifier runs server-side by default for:
- Claude API users
- Claude Enterprise accounts
- Bedrock deployments
- Vertex AI deployments
- Azure AI Foundry deployments
- Claude apps gateway deployments
The server-side classifier does not charge for classifier overhead. When server-side classification falls back to local (e.g., network unavailability), Claude Code emits a warning so you know the billed path was taken.
For local-only deployments or deployments that cannot reach Anthropic’s servers: set CLAUDE_CODE_AUTO_MODE_SERVER=0 in the environment to retain local-classifier behavior.
The /status panel now includes an Auto mode server row so you can confirm at a glance which classification mode is active.
Try It
- Check
/statusafter upgrading to v2.1.278+ — the newAuto mode serverrow confirms server-side classification is active. - If you run air-gapped or proxy-only deployments: test whether server-side classifier calls succeed through your proxy before upgrading. If not, set
CLAUDE_CODE_AUTO_MODE_SERVER=0. - Monitor for “billed fallback” warnings in the first sessions after upgrading, especially in Bedrock/Vertex/Foundry deployments where the routing path is more complex.
Open Questions
- What triggers a “billed fallback”? The changelog entry says Claude Code warns on billed fallback, but does not specify which conditions (network timeout, classification error, rate limit) cause it.
- Does this affect local/non-API deployments? The entry lists Claude API, Enterprise, Bedrock, Vertex, Foundry, and gateways. It is not stated whether
claude -psessions with a local model or offline deployments are affected.
Related
- What’s New — Week 37 — the preceding digest;
claude plugin eval, AGENTS.md, send-now key, claude.ai skills sync. - What’s New — Week 39 — Claude Opus 5.5 launch; default model changes; /doctor prompt-audit.
- CLI Reference —
CLAUDE_CODE_AUTO_MODE_SERVERenv var;/statuspanel.