Session recap (away summary) ignores ANTHROPIC_DEFAULT_HAIKU_MODEL and bills the session's main model

Status Open
Reported on v2.1.227
Maintainer reply None cached
Activity 0 comments · opened Aug 11, 2026

Summary

ANTHROPIC_DEFAULT_HAIKU_MODEL is documented as the model used for background
functionality, but the automatic session recap shown on idle return does not
honor it. The recap request goes to the session's main model instead.

Session-title generation, another background feature, does honor the variable in
the same session, so the variable is being read and the alias resolves. The recap
just doesn't use it.

Expected

With ANTHROPIC_DEFAULT_HAIKU_MODEL set, background requests (session title,
idle recap) resolve to the configured Haiku-class model regardless of the
session's main model.

Actual

One session, v2.1.227, ANTHROPIC_DEFAULT_HAIKU_MODEL set to a Haiku model ID in
settings, main model set to Opus at session start. Request-level telemetry for
that session, keyed by the feature that issued each call:

| Feature | Model | Prompt / output tokens | Cost |
| :-- | :-- | :-- | :-- |
| generate_session_title | configured Haiku model | 623 / 18 | $0.0007 |
| repl_main_thread (42 calls) | Opus | — | $3.81 |
| away_summary (idle recap) | Opus | 148,589 / 73 | $0.09 |

The recap arrived in the transcript 1.6s after that request. Of 44 requests in the
session, exactly one ran on the configured Haiku model: the title.

The recap is the costly one because it carries the whole conversation as context
and fires on every idle return, so a session left open across a workday pays it
repeatedly for two sentences nobody asked for.

Reproduce

  1. Set ANTHROPIC_DEFAULT_HAIKU_MODEL to a Haiku model ID in settings.
  2. Start a session, set the main model to Opus, send a prompt that builds real

context.

  1. Leave it idle long enough for the automatic recap to fire, then return.
  2. Check request-level telemetry or your gateway logs. The title request is on

Haiku; the recap request is on the main model.

Environment

  • Claude Code 2.1.227, macOS, CLI entrypoint
  • ANTHROPIC_BASE_URL pointed at an internal LLM gateway serving both the Haiku

alias and the main model

Notes

If routing the recap to the main model is intentional, please say so in the docs.
ANTHROPIC_DEFAULT_HAIKU_MODEL currently reads as covering all background
functionality, and the recap is the costliest of those features by a wide margin.
Documenting CLAUDE_CODE_ENABLE_AWAY_SUMMARY would also help. It isn't in the
env-vars reference today, so the only reliable way to control this cost is
undocumented.

Related: #58870 (docs gap on background-functionality model fallback for custom
gateways).

View original on GitHub ↗