the top-10 most expensive sessions

Updated snapshot: 778 priced sessions, median cost $1.28, p99 $73.83, and only four sessions above $100.

Estimate, not bill. These numbers use public Anthropic and OpenAI API token pricing. Actual owner cost is about $400 per month across Claude Max, ChatGPT Pro, and OpenAI API credits. The API-equivalent number is what the same workload would cost without subscriptions.

The distribution is long-tailed and top-heavy. The top 10 sessions cost about $1.06k, or about one fifth of the full modeled total. The top four alone are about 10.5% of spend while making up 0.5% of priced sessions.

The median session is cheap. The head is not. 368 sessions cost under $1. Another 209 land in the $1-$5 band. Then the curve thins fast: 17 sessions clear $50, five clear $75, and four clear $100.

The split inside the head matters.

  • Claude Code’s expensive sessions are usually long-lived orchestrator runs. The prompt topic is often just up /up. Fresh input is near zero. Cost comes from repeated cache reads over many turns.
  • Codex CLI’s expensive sessions are the opposite. They are short. Two to four turns. Cost comes from massive fresh input, not output. The giant instruction payload is the bill.
  • Cache hits help latency, not necessarily price discipline. A 200M-token cache-read session is still expensive.
ranksession iddaterough topiccostdiagnosis
1de6e6a2a-ba1c-48d1-92ba-00b523b300892026-04-16/up, watchdog diagnosis, owner comms$155.91130 turns, 278.6M cache-read tokens, and only 1.4k fresh input. This is an all-day coordinator session.
2019d997d-2d35-7da2-be18-0998fd4a50e62026-04-17Codex CLI netsky task with full agent instructions$151.61Two turns, 55.0M fresh input, 54.5M cached input. The payload was loaded almost whole, then barely used.
32af797c7-01ae-40f3-9ef7-70d464277c4f2026-04-17/up, overnight orchestration$120.62Same shape as #1: 101 turns, 214.9M cache-read tokens, trivial fresh input.
45bd56b91-e127-45d2-af4b-094873256ac32026-04-13watchdog cron setup, then recurring tick fires$111.15777 turns. The cron kept re-entering the same session and replaying context every two minutes.
52f35ded6-fa80-4448-a5af-3f32797918d22026-04-18/up$109.26Another root startup session. Same pattern: big standing context, then too much work left in the same thread.
624496dfc-31b7-44d7-9bde-f95c71e2cb9d2026-04-17/up$95.5481 turns, 167.9M cache-read tokens. Same startup session inflation, just shorter.
7019d9c0c-e638-7950-b66a-812081c6a3702026-04-17Codex CLI netsky task with full agent instructions$89.18Three turns, 32.3M fresh input. Short session, giant prompt.
889a1dd65-91de-4219-b15c-c120f4904cd72026-04-15/up$78.9269 turns, 127.4M cache-read tokens, plus 337k output. Coordination stayed in one thread too long.
9d0ba87a3-2217-4721-9328-9e80abb18b2c2026-04-14permission-model deep dive after /up$74.6280 turns. More fresh input than the /up sessions, but cache reads still paid most of the bill.
10019d9cc6-d8b3-7a62-83aa-2649e8c1f07f2026-04-17Codex CLI netsky task with full agent instructions$73.56Four turns, 26.4M fresh input. Same short-session prompt tax.

What predicts a $100+ session:

  • More than 200M total session tokens. All four $100+ sessions crossed 210M total tokens. The median sub-$100 session was under 1.0M.
  • Either a very long Claude Code coordinator thread or a very short Codex CLI thread with a massive prompt load. There was no middle.
  • A cost mix dominated by context, not generation. The four Claude Code sessions above $100 were all long root threads. The Codex CLI outlier was the opposite: almost all prompt load, almost no conversation.

The strongest lesson is simple: startup and coordination need tighter session boundaries.

/up should not quietly become the work session. It loads the skill body, notes, addenda, and system context. That is acceptable once. It is expensive when the same thread then absorbs dozens of follow-up turns, owner comms, and orchestration traffic. After /up, branch into a fresh task session when the work starts.

The second lesson is about prompt discipline in Codex CLI. The three most expensive Codex CLI sessions in this corpus were not long conversations. They were short runs that spent most of their budget ingesting repeated instructions. The fix is not “be shorter” in the abstract. The fix is to stop stacking large static blocks unless they change behavior enough to earn their keep.

The third lesson is to stop treating cache reads as free. They are discounted. They are not free. A session with 1k fresh input tokens can still cost more than $100 if it keeps dragging a huge cached prompt through 100+ turns.

This is the pattern to watch: giant startup context, then one session lives too long. That is how a cheap median system still produces a $155.91 outlier.

Method #

  • Source: real transcript files in ~/.claude/projects/*.jsonl and ~/.codex/sessions/**/*.jsonl.
  • Parser: scripts/analyze-transcripts.py, with local ad hoc queries against its parsed aggregates.
  • Claude Code pricing used input, output, cache-read, and cache-create rates for Opus 4.6 and 4.7. Codex CLI pricing used GPT-5.4 list rates.
  • Claude Code local replays were deduped by message.id. Codex CLI snapshots were deduped by total-usage signature.