Skip to content

Usage Evidence

Jump to bottom
xsyetopz edited this page Oct 13, 2026 · 2 revisions

Usage Evidence

This page gives the facts about the plans, the prices, the models, and the usage of dotclaude 1.0.0. It also gives the measurements of one heavy week and what they mean. Read it to learn what drives the usage, and which values are measured. The source of each number is named in its row or line. Source labels are official, binary, measured, reported, and inference.

Summary

  • One Max 20x week reached 100% of the weekly limit at $1,854 (measured, see the measured week).
  • Three things drive the usage: the context per turn, the subagent fan-out, and the general-purpose agent. The model choice is not one of them.
  • Turns set the cost, not tool calls.
  • dotclaude 1.0.0 does not detect the plan for its policy. Setup picks an effort and a subagent cap from the account tier only.
  • dotclaude 1.0.0 does not use opusplan.

Models and Efforts of 1.0.0

Each agent sets its model and effort in its frontmatter. The source is src/plugins/dotclaude/slices/agents/agents/*.md.

Agent Model Effort
planner claude-opus-5-5 high
reviewer claude-opus-5-5 high
debugger claude-sonnet-5-5 high
researcher claude-sonnet-5-5 medium
implementer claude-sonnet-5-5 medium
investigator claude-haiku-5-5 medium

The settings snippet sets these values, from src/plugins/dotclaude/slices/prompt/slice.ts.

  • model is claude-opus-5-5.
  • CLAUDE_CODE_SUBAGENT_MODEL is claude-sonnet-5-5. An agent file that names its own model overrides it.
  • maxEffortLevel is max. The snippet sets no effortLevel.
  • promptCacheTtl is 5m. The snippet sets no subagentPromptCacheTtl. It also sets no advisorModel.
  • availableModels lists claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5, and claude-fable-5-1.
  • The variable CLAUDE_CODE_EFFORT_LEVEL overrides the effort of the agent files. Its name is in src/plugins/dotclaude/slices/setup/domain/facts.ts.

Account Tiers and Setup Picks

Setup reads claude auth status --json and the rate limit tier. It then writes an effort and a subagent cap for each account. The source is src/plugins/dotclaude/slices/setup/domain/tiers.ts.

Account effortLevel CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS
Pro high SUBAGENTS_PRO
Max 5x xhigh SUBAGENTS_MAX_5X
Max 20x xhigh SUBAGENTS_MAX_20X
API key high SUBAGENTS_PRO

The subagent caps are in src/plugins/dotclaude/slices/setup/limits.ts. Setup does not set the plan mode agents or the usage warnings. Claude Code scales them by tier. The account with no login gets no pick.

Prices

The API list prices are per million tokens. They were checked on 2026-10-13 (official). Source: https://platform.claude.com/docs/en/about-claude/pricing.

Model Base input 5-minute cache write Cache read Output
Fable 5.1 $10 $12.50 $0.25 $50
Opus 5.5 $4 $5 $0.20 $20
Sonnet 5.5 $2 $2.50 $0.10 $10
Haiku 5.5 (prompts up to 100,000 tokens) $0.10 $0.125 $0.01 $0.50
Haiku 5.5 (prompts over 100,000 tokens) $0.50 $0.625 $0.05 $2.50
  • The Haiku 5.5 price rises above 100,000 prompt tokens. The investigator agent is the only Haiku 5.5 agent of 1.0.0.
  • The Sonnet 5.5 cache read is $0.10 per million tokens on this page. An older text of this wiki gave $0.20. That value is stale.
  • Anthropic publishes no subscription multiplier.

Fable Limit

Fable draws from the regular weekly limit. It is not an extra pool. These facts are from the Anthropic help pages, as the old text of this page cited them. They are not checked again in this run.

  • On Max and on Team or Enterprise premium seats, Fable can use up to 50% of the weekly limit.
  • On Pro and on standard Team or Enterprise seats, Fable runs on usage credits.
  • The binary labels the seven_day_overage_included limit as the Fable limit (binary).

A Fable weekly figure of 0% does not give free headroom. Each model has at most the remaining share of the weekly limit.

Cache Facts

  • An effort change keeps the prompt cache. A model change rebuilds the prompt cache, because the cache is kept for each model. This is official, from the prompt caching page.
  • The cache of the main conversation lives 1 hour on a subscription within its limits. It lives 5 minutes on usage credits or an API key. Subagents use 5 minutes (official, as the old text of this page cited it, not checked again in this run).
  • A prompt after the TTL, and a /compact after it, write the whole context again. dotclaude 1.0.0 sets promptCacheTtl to 5m.
  • From Claude Code 2.1.287, the SessionStart input for resume and fork has prompt_cache_likely_expired (binary, not checked again in this run).

Effort

  • Claude Code defaults Opus 5.5 and Sonnet 5.5 to medium (official, not checked again in this run).
  • dotclaude does not set a top-level effortLevel in the snippet. The setup picks of the account tiers set it, as shown above.
  • No hook event fires on an effort change (binary).

Opusplan

dotclaude 1.0.0 does not use opusplan. Nothing in src/ names it. opusplan switches the model when plan mode starts and ends. A model switch loses the prompt cache (official).

The Measured Week

One Max 20x week (2026-09-21 to 09-28) reached 100% of the weekly limit. The figures come from a scan of the local transcripts (measured). The repository does not hold that scan. The transcripts start on 09-25, so the figures are a lower bound.

The week ran under a 0.x config, not under 1.0.0. Most of the week ran before 0.5.0 moved implementer to Sonnet 5, so the implementer row ran mostly on Opus 5.5. The source is the 0.26 text of this page, in git commit af3071f4 (wiki/Usage-Evidence.md, line 22). The figures show the 0.x config, and they do not describe the 1.0.0 agents of the table above.

Measure Value
Total $1,854 (Opus 5.5 98.7%)
Cost parts cache reads 64.7%, cache writes 28.1%, output 7.2%
Main conversation 44.9%, context p50 369k, p90 807k tokens
implementer 30.7%, 155 runs, median 68 calls
general-purpose 17.6%, 87 runs, median 51 calls
Calls with context over 150k 74.7% of cost
Subagent cache writes (5-minute TTL) 71M tokens, about $355
First cache write of a fresh subagent p50 13.9k, p90 24.8k tokens
Main turns that background agents started 501 of 861, about $385

Three things drive the usage.

  1. Context per turn. Each turn reads the whole context again.
  2. Subagent fan-out. Each fresh subagent writes about 14k tokens to the cache before its first tool call. Its 5-minute cache expires while it waits on long builds.
  3. general-purpose in place of dotclaude agents. The settings snippet denies Agent(general-purpose) in src/plugins/dotclaude/slices/prompt/slice.ts. The denial is a setting. It is not a proof that the habit stopped.

The Delegation Share

just usage runs plugins/dotclaude-sessions/bin/dotclaude-usage. It repeats the scan on any machine. It reports the sessions by entrypoint, the skill calls, the delegation share, and the latest subagent runs.

  • measured, 2026-10-03, 7 days: 976 subagent runs in 25,578 main turns, which is 3.8 per 100 turns.
  • measured, 2026-10-03: tool-result tokens per main session, median 19,169, p90 167,675 (253 sessions).

Turns, Not Tool Calls

Each model turn reads the context again. So the number of turns sets the cost.

  • Independent tool calls in one response take one turn. The same calls in sequence take one turn each.
  • Claude Code has no general tool that runs several tools in one call (binary).

Community Claims

  • True: /compact reads the whole conversation to write its summary. On a large context it is an expensive request (official). /clear costs nothing.
  • True: MEMORY.md loads at every start, up to its first 200 lines or 25 KB (official). Topic files in the memory folder do not load at start.
  • True: a short CLAUDE.md saves usage. It loads at every start, and in every subagent that does not set omitClaudeMd.
  • True: subagents keep large reads out of the main context.
  • False: old .jsonl transcripts in ~/.claude/projects/ do not go into the context. Only /resume loads one. The settings snippet does not set cleanupPeriodDays. See settings snippet keys.

Tools That dotclaude Does Not Use

Tool Reason
RTK A JetBrains measurement of 7.6% more cost (reported).
Graphify It fails on large repositories (reported).
grepai It needs embeddings from Ollama or OpenAI, and it has no measurement.

Measuring a Session

just measure runs tools/measure.ts. It reads a session transcript and prints where the context went. It reports the first request of the session, each subagent, and the repeat reads of the same file. The measures that led to the context bounds are in context and handoff.

Related Pages

Home | README | CHANGELOG | Issues

Clone this wiki locally