TokenSave Data · Dataset

Claude Code's Explore subagent stopped using Haiku in 2.1.198, and two settings cut the cost by 60%: 205 runs

Up to Claude Code 2.1.197, the built-in Explore subagent ran on Haiku. From 2.1.198 it runs on whatever model the main conversation uses. With Opus as the main model, that took a delegated code search from $0.18 to $0.37 per run (+105%). On 2.1.318 with Opus 5.5, setting CLAUDE_CODE_SUBAGENT_MODEL=haiku together with CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 put Explore back on Haiku and cut the cost per run by 60%, with the same share of answers found.

Measured 2026-11-08 · 205 runs · 26 Claude Code versions · License CC BY 4.0

DOI 10.5281/zenodo.24048094 · Also on Hugging Face · Kaggle · GitHub

Key findings

• The switch is 2.1.198. With the main model fixed to Opus 4.5, Explore ran on Haiku 4.5 in every run from 2.1.0 to 2.1.197 and on Opus 4.5 in every run from 2.1.198 to 2.1.318. A February 2026 post saying Explore is “hard-coded to Haiku” was right at the time.
• It doubled the cost of a delegated search. Same prompt and questions: $0.18 per run while Explore was Haiku (40 runs), $0.37 after (44 runs). Share of right answers found: 89% before, 94% after, on 2 questions per version, so read that gap as small.
• CLAUDE_CODE_SUBAGENT_MODEL alone does nothing for Explore. On 2.1.318 with Opus 5.5, setting it to claude-haiku-5-5 left Explore on Opus 5.5 in all 36 runs: $0.37 per run vs $0.38 by default.
• Add CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 and it works. Explore ran on Haiku 5.5 in all 36 runs. Answers found: 100% vs 100% by default on the same 12 questions × 3 runs. Cost per run: $0.38 → $0.15 (-60%).
• This matches the docs. The sub-agents page says Explore runs on the main conversation’s model, that CLAUDE_CODE_SUBAGENT_MODEL alone does not change Explore, and that the FORCE flag (v2.1.257+) applies it to every subagent. It also says a project subagent named Explore with model: haiku overrides the built-in; we did not test that route. What the data adds is when the default changed and what each choice costs.

How to get Haiku back for Explore

Put both lines in your environment or in settings.json under env. The FORCE flag needs Claude Code 2.1.257 or later, and it applies the model to every subagent, not only Explore.

CLAUDE_CODE_SUBAGENT_MODEL=haiku
CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1

2.1.318 with Opus 5.5: three settings

Same prompt (“use subagents (the Explore agent)”) on dataset 27’s 12 code-search questions, 3 runs each. Found = share of the answer key named in the answer, mean per question.

Version by version

Main model Opus 4.5 (Opus 5.5 is refused below 2.1.280). Questions M1 and U1 from dataset 27, 2 runs each; versions with 1 run are bisection probes on M1.

Method

• Pre-registered on 2026-11-07: PREREG.md, with two dated amendments.
• Prompt and scoring from dataset 27: a “find every place” question plus an instruction to use the Explore subagent, scored against a regex-and-parser answer key.
• Agent: headless claude -p, read and search tools only, no MCP servers, Windows 11, a separate config directory. The Explore model is read from the subagent’s own messages and checked against the per-model cost split.
• Incidents: 21 runs hit the subscription’s 5-hour limit and were re-run. The first bisection used a regex that never matched a Haiku entry and wrongly put the switch at 2.1.176; the raw logs showed Haiku through 2.1.186, and the bisection was re-run (amendment 2).
• Cost is Claude Code’s own per-run figure at API prices; the runs were made on a subscription.

Limitations

• Two questions per version: the history shows which model ran, not how good it was.
• One machine, one account, one prompt wording. Defaults may change again.

How to cite

Free to use, quote and republish with attribution under CC BY 4.0. Please link to this page.

BibTeX

@misc{tokensave_claude_subagent_model_by_version_2026,
  author    = {{TokenSave}},
  title     = {Claude Code's Explore subagent stopped using Haiku in 2.1.198, and two settings cut the cost by 60\%: 205 runs},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.24048094},
  url       = {https://doi.org/10.5281/zenodo.24048094},
  note      = {Dataset, CC BY 4.0}
}

Embed the chart

Related

• Subagent effort in Claude Code: 288 runs (dataset 27)
• Edge-case prompts and effort: 504 runs (dataset 26)
• Estimate your own Claude Code or Codex bill

Questions or corrections: contact@jonhisking.com