Claude Code v2.1.278 Defaults Auto Mode to Server-Side Classifier
Claude Code v2.1.278 changes how auto mode picks its classifier. For Claude API and Enterprise users, plus anyone running on Bedrock, Vertex, Foundry, or gateways, auto mode now defaults to the server-side classifier. The practical upside: that path does not charge for classifier overhead.
What changed
- Server-side classifier is now the default for Claude API users, Enterprise users, and deployments on Bedrock, Vertex, Foundry, and gateways.
- No billed classifier overhead on that server-side path.
- Opt-out flag:
CLAUDE_CODE_AUTO_MODE_SERVER=0disables the server-side classifier on Bedrock, Vertex, Foundry, and gateways. - Billed fallback warning: the CLI warns you when it falls back to a path that bills for classifier work.
Checking which classifier your session uses
v2.1.278 adds an Auto mode server row to the /status output. Run /status in a session and you can see whether that session's auto mode classifier is running on the server or falling back locally. Useful when you're debugging cost or behavior differences across environments.
Opting out
If you're on Bedrock, Vertex, Foundry, or a gateway and want to keep the client-side classifier, set the environment variable before launching:
CLAUDE_CODE_AUTO_MODE_SERVER=0 claude
Per the release notes, this opt-out applies to those four deployment paths specifically; Claude API and Enterprise users get the server-side classifier by default.
Billing details
The release points to Anthropic's docs for the full billing behavior of the auto mode classifier: https://code.claude.com/docs/en/auto-mode-classifier-billing. If you're tracking per-token or per-request costs on a gateway or cloud marketplace deployment, read that page before flipping the env var — the whole point of the default change is that the server-side path is the cheaper one.
Who this matters for
Anyone running Claude Code auto mode through Bedrock, Vertex, Foundry, or a self-managed gateway, and Enterprise accounts watching classifier line items on their bill.
📖 Read the full source: GitHub Claude-Code
👀 See Also
The Frontier AI Race is Over: Networks of Smaller Models Beat Centralized AI on Cost and Capability
Networks of smaller AI models now outperform every frontier AI system on speed, accuracy, and cost. The article argues that centralized AI companies cannot regain the lead due to the "Hydra Effect" — ensembling cheaper models recursively beats any single model.
AI Is Too Expensive: Hyperscalers Need $3 Trillion to Break Even
Hyperscalers have invested over $800B in AI capex, with $1T more planned for 2027. Microsoft alone spent ~$100B on OpenAI infrastructure, yet AI revenue covers only ~20% of its capex.
Hy3 LLM Tops OpenRouter Rankings: Cheapest Model or Something Else?
Hy3 preview, a Tencent open-source LLM, surged to the top of OpenRouter's model rankings by token usage, surpassing Claude and DeepSeek V4 Flash. Priced at $0.066/1M input tokens, it's the cheapest major model, but benchmarks show quality far below leaders.
Claude Code Performance Regression Diagnosed: Configuration, Not Model Intelligence
Anthropic's postmortem reveals Claude Code's performance drop was caused by three product changes — default reasoning effort, session caching bug, and prompt-verbosity — not model degradation. Rollback restored performance.