<!-- llms-explorer concept facts · https://llms-explorer.com/tree/claude-code-helper-model-request-to-custom-base/ · pack 2026-10-05 · ~878 tokens -->

# Claude Code helper-model request to custom base URL

> ANTHROPIC_DEFAULT_HAIKU_MODEL sets what the haiku alias resolves to and is also the model used for background functionality; pointing it at the local model ID sends helper requests to the same served model.

Parent: [Mac local LLMs: Agent clients, context and compaction](https://llms-explorer.com/tree/mac-local-llms-agent-clients-context-and-compaction/) · 1 facets · 17 facts · page: https://llms-explorer.com/tree/claude-code-helper-model-request-to-custom-base/

## Facts

- ANTHROPIC_DEFAULT_HAIKU_MODEL sets what the haiku alias resolves to and is also the model used for background functionality; pointing it at the local model ID sends helper requests to the same served model. — source: `asserted`
- ANTHROPIC_SMALL_FAST_MODEL is the deprecated older name for the same Haiku-class background model. — source: `asserted`
- Documented background work, even when idle: summarizing previous conversations for claude --resume; the costs page says it uses a small number of tokens. — source: `asserted`
- With prompt suggestions on, a short request goes to the session's own model after each response and reuses the conversation prompt cache, so it is a main-model request, not a helper-model one. — source: `asserted`
- CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 turns off auto-updates, telemetry, error reporting, feedback and availability checks; the docs do not list helper-model calls among them. — source: `asserted`
- ANTHROPIC_SMALL_FAST_MODEL is marked deprecated in the current env-vars page; ANTHROPIC_DEFAULT_HAIKU_MODEL replaces it. — source: `asserted`
- On a single-slot local server the helper request carries a different prefix (own system prompt, 27 tools) so it can evict the main session's cached prefix; this builds on the existing measurement. [inferred] — source: `asserted`
- If the haiku alias is left at its default, a server that serves one model ID returns 404 for it (existing LM Studio report). — source: `asserted`
- Whether a pinned local ID removes the helper request or only redirects it; not measured here. — source: `asserted`
- Which helper requests (title, summary, command-prefix check) fire per session and whether any can be switched off individually. — source: `asserted`
- ANTHROPIC_DEFAULT_HAIKU_MODEL sets the haiku alias target and is also used for background functionality — [source](https://code.claude.com/docs/en/env-vars)
- ANTHROPIC_SMALL_FAST_MODEL is deprecated and names the Haiku-class model for background tasks — [source](https://code.claude.com/docs/en/env-vars)
- Claude Code runs background jobs that summarize previous conversations for claude --resume, using a small number of tokens even when idle — [source](https://code.claude.com/docs/en/costs)
- Prompt suggestions send a short request to the session's own model and reuse its prompt cache — [source](https://code.claude.com/docs/en/costs)
- CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC disables updates, telemetry, error reporting and feedback traffic — [source](https://code.claude.com/docs/en/env-vars)
- Under an LLM gateway set through ANTHROPIC_BASE_URL, a listed model ID with no built-in picker row appears as its own row, and ANTHROPIC_DEFAULT_*_MODEL cannot redirect an allowed alias outside availableModels — [source](https://code.claude.com/docs/en/model-config)
- Pinning ANTHROPIC_DEFAULT_HAIKU_MODEL to the served local model ID is the documented way to keep helper requests on a model the server has — source: `asserted`
