Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

While there's a ton of other issues here the one I want to address is the increase in output tokens starting with version 2.1.98 (2x increase in output tokens) and increasing from there (2.1.101+) to 3x+ in output tokens.

Environment

• Claude Code versions tested (pinned via ~/.local/share/claude/versions/<ver>, DISABLE_AUTOUPDATER=1): 2.1.96, 2.1.98, 2.1.101, 2.1.107
• Model: claude-sonnet-4-6 (same in all arms)
• Mode: --print --permission-mode acceptEdits (non-interactive)
• OS: macOS 15 (Darwin 25.3.0)
• Installer: native (version file is the executable)

TL;DR

Running identical prompts against identical tasks on identical models, Claude Code's per-assistant-turn output token usage jumps in two discrete steps inside the v2.1.96 → v2.1.101 range, then plateaus:

Both arms reproduce in ~60 seconds in an empty directory with public data. This is not a session-length, tooling, or workflow artefact — the prompt is byte-identical and the task is a one-shot CSV-to-HTML report.

Reproducer

Minimal harness attached in cc-ab/ — empty working directory, one public CSV (Palmer Penguins, CC0), one prompt that asks for a self-contained HTML report with three charts. Four run directories, each pinning one Claude Code version:

# from the harness root
cd run-2.1.96 && bash run.sh     # ~60s
# wait ≥ 5 min for prompt cache TTL to expire
cd run-2.1.98 && bash run.sh     # ~60s
# wait ≥ 5 min
cd run-2.1.101 && bash run.sh    # ~60s
# wait ≥ 5 min
cd run-2.1.107 && bash run.sh    # ~60s

# compare
cd .. && python3 analyze.py

Each run.sh invokes:

"$HOME/.local/share/claude/versions/$CC_VERSION" \
  --model claude-sonnet-4-6 \
  --print \
  --permission-mode acceptEdits \
  "$(cat ../prompt.md)"

Full metrics

Output-per-assistant-message is the cleanest signal — it normalises for session length and prompt count. It doubled at 2.1.98 on the same workload.

Cache-write is nearly flat across arms (1.00× / 1.06× / 1.13×) because all runs are back-to-back with no idle gaps that would expire the 5-minute prompt cache TTL. Earlier real-workload runs over a 6-hour session showed a 2.04× cache-write delta that vanished here, so cache-writes are partly an idle-gap artefact — output tokens are the load-bearing signal.

Session JSONLs (attached, email-redacted, otherwise intact)

Each JSONL contains: all per-message usage metrics, version, model, all tool-use blocks, MCP tool advertisements (deferred_tools_delta), MCP instruction blocks, skill listings, and SessionStart hook output. Only the reporter's email address is redacted — everything else is intact to help triage.

Matched pairs of report.html (the actual deliverables each run produced) are also attached under run-*/report.html and are substantially similar in output — i.e. v2.1.107 is not producing a meaningfully larger artefact to justify 3.7× the output tokens.

What this rules out

• Prompt / work-mix — prompts are byte-identical across arms.
• Model — all arms run claude-sonnet-4-6.
• User behaviour / machine — same user, same machine, back-to-back runs.
• Project size / complexity — minimal reproducer runs in an empty directory with a 15 KB CSV. The regression still shows up.
• Workflow shape — no sub-agents invoked, no orchestration, no resume, no compaction. Single turn with two file writes.
• More actual work done — v2.1.98 produces the same session shape as v2.1.96 (5 asst messages, 2 tool uses, identical user turns) yet doubles output tokens. v2.1.107 produces one extra Read call, same final artefact.

What's left as the variable

Claude Code itself, between 2.1.96 and 2.1.101 — and specifically two separate changes in that range, each amplifying per-turn output.

Changelog context (from Anthropic's public changelog):

• 2.1.98 — sub-agent partial-progress now reports up to the parent session (attributed to parent output tokens); live /agents layout; Monitor tool added. This is the first-jump inflection in the harness. But the harness invokes zero sub-agents, which suggests either (a) the sub-agent attribution path also runs in no-sub-agent sessions, or (b) another 2.1.98 change is responsible.
• 2.1.99 / 2.1.100 — would need Anthropic to identify which behaviour-affecting change in this range caused the second step.
• 2.1.101 — --resume preserves more context on large sessions; sub-agents inherit MCP tools. By this version, per-turn cost has reached the full ~2.7× plateau.

2.1.105 (skill description cap 250 → 1,536 chars, 6×) is a separate known cost amplifier but can't explain this harness — full regression is visible at 2.1.101, which predates 2.1.105.

Ask

1. Confirm whether the per-turn output increase at 2.1.98 and the further increase by 2.1.101 are intentional at this cost profile. User-visible result: sessions that used to run cheaply now cost ~2.7× more in output tokens for the same work.
2. Identify the changelog entries responsible for each of the two step changes (2.1.96 → 2.1.98 and 2.1.98 → 2.1.101). If it's sub-agent partial-progress attribution, please clarify why it affects sessions that invoke zero sub-agents.
3. Expose knobs for the highest-impact drivers:
   • An opt-out (or cap) on sub-agent output-token attribution when the sub-agent output isn't consumed by the parent.
   • An opt-out (or cap) on the 2.1.105 skill description expansion for users with many installed skills.
   • A higher or user-configurable context threshold so mid-session compactions (each ~5–15 K output tokens) are less frequent.

What Should Happen?

Different versions running the same task should have, roughly, the same token usage.

Error Messages/Logs

Steps to Reproduce

1. Install multiple versions of claude using curl -fsSL https://claude.ai/install.sh | bash -s 2.1.98 where "2.1.98" if the version number to install recommend 2.1.96 (last good version, 2.1.98 (first broken version) and 2.1.101 (most broken version)
2. Create run.sh with the following:

#!/usr/bin/env bash
# A/B harness — arm A: Claude Code v2.1.96
# Runs the penguins-report prompt non-interactively against the pinned version.
# Session JSONL lands at ~/.claude/projects/<cwd-slug>/<uuid>.jsonl
# (Claude Code replaces '/' and '.' in the cwd with '-' to form the slug.)
set -euo pipefail

# change for version you want to run,  version must be installed. 
CC_VERSION="2.1.96" 
CC_BIN="$HOME/.local/share/claude/versions/$CC_VERSION"

if [[ ! -x "$CC_BIN" ]]; then
  echo "ERROR: Claude Code $CC_VERSION not found at $CC_BIN" >&2
  echo "Install with: npm install -g @anthropic-ai/claude-code@$CC_VERSION" >&2
  exit 1
fi

HERE="$(cd "$(dirname "$0")" && pwd)"
cd "$HERE"

# Ensure dataset is present (self-contained run dir)
if [[ ! -f data/penguins.csv ]]; then
  mkdir -p data
  cp ../data/penguins.csv data/penguins.csv
fi

# Prompt lives one level up, shared across arms
PROMPT_FILE="../prompt.md"

export DISABLE_AUTOUPDATER=1
export CLAUDE_CODE_VERSION_PINNED="$CC_VERSION"

echo "=== Claude Code $CC_VERSION — penguins report ==="
echo "Working dir: $HERE"
echo "Prompt:      $PROMPT_FILE"
echo "Started:     $(date -Iseconds)"
echo

"$CC_BIN" \
  --model claude-sonnet-4-6 \
  --print \
  --permission-mode acceptEdits \
  "$(cat "$PROMPT_FILE")"

echo
echo "Finished: $(date -Iseconds)"
echo "Session JSONL: ~/.claude/projects/$(pwd | sed 's|/|-|g')/*.jsonl"

Next to that file you will need a data directory with the attached penguins.csv file.

1. Run bash run.sh

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.96

Claude Code Version

2.1.107

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

Full A/B test harness attached.

cc-ab-clean.zip - ready to test - no results yet cc-ab-with-results.zip - with tests and analysis.