Skip to content

Claude Opus 5.5 on Analyze five months of AI coding-agent session logs

AClaude Opus 5.5 · Claude Code 2.1.299 (headless -p) · cloud · 2026-10-14 · by @bobsiopener

Trust 80%Fully autonomoussuccess · first try
Estimated cost
$0.12
Wall time
21s
Human step-ins
0
Outcome
Pass

Compare with other models on this task →

Summary

Exact prompt

sessions.csv logs 5 months of AI coding-agent sessions. Write answers.json with exactly these keys: "total_sessions" (int), "total_prs" (int), "total_api_equiv_usd" (number rounded to 2 decimals), "busiest_month_by_sessions" (string like "2026-07"), "median_minutes" (number), "usd_per_pr_by_tool" (object tool -> total api_equiv_usd / total prs_opened for that tool, rounded to 2 decimals). Then write REPORT.md with a 5-bullet summary.

Task prompt given to the agent (headless, offline, synthetic fixture inputs): sessions.csv logs 5 months of AI coding-agent sessions. Write answers.json with exactly these keys: "total_sessions" (int), "total_prs" (int), "total_api_equiv_usd" (number rounded to 2 decimals), "busiest_month_by_sessions" (string like "2026-07"), "median_minutes" (number), "usd_per_pr_by_tool" (object tool -> total api_equiv_usd / total prs_opened for that tool, rounded to 2 decimals). Then write REPORT.md with a 5-bullet summary. Inspired by https://x.com/quinn_leng/status/2085828592127336541 (unverified; not checked with the author). This is an independent re-run on a reproducible synthetic task; it does not report or reproduce the original post's numbers. Cost shown is a list-price API-equivalent estimate (run on a Claude Pro subscription, not billed).

Success criteria: All automated rubric checks pass: total_sessions correct; total_prs correct; total_api_equiv_usd correct; busiest_month_by_sessions correct; median_minutes correct; usd_per_pr_by_tool correct; REPORT.md written.

Setup notes: Cost is an ESTIMATE: Claude Code's list-price API-equivalent figure. The run used a Claude Pro subscription, so nothing was billed. input_tokens = all input incl. cache-creation and cache-read tokens; cached_tokens = cache reads (subset of input). Claude Code's cost bills cache writes at the 1h cache-write premium, so it can differ from a flat input-price recompute. Web disabled; tools Read/Write/Edit/Glob/Grep/Bash; single run; scored automatically (7/7 checks).

Estimated cost List-price estimate; not actually billed.

Metrics

Estimated cost
$0.12
Wall time
21s
Human step-ins
0
Turns
5
Model usage
ModelInput tokOutput tokCost
AClaude Opus 5.5anthropic/claude-opus-5.593,5262,023$0.12

Reproduce this run

  1. Use the use case prompt on the use case page.
  2. Same setup: Claude Opus 5.5 (anthropic/claude-opus-5.5) via Claude Code 2.1.299 (headless -p), cloud.
  3. It counts as a success only if it meets every success criterion: All automated rubric checks pass: total_sessions correct; total_prs correct; total_api_equiv_usd correct; busiest_month_by_sessions correct; median_minutes correct; usd_per_pr_by_tool correct; REPORT.md written.
  4. Log model, tokens, wall time, human step-ins and the evidence log, then submit it as a reproduction. A different model? Submit a head-to-head rerun.

How runs are scored: Methodology.

Was this useful?

⇄ Rerun on another model

Automated checks

Trust 80% comes from these checks (how Trust % works).

Evidence log
{
 "duration_api_ms": 19643,
 "stop_reason": "end_turn",
 "session_id": "[REDACTED]",
 "total_cost_usd": 0.1201134,
 "usage": {
  "input_tokens": 10,
  "cache_creation_input_tokens": 7809,
  "cache_read_input_tokens": 85707,
  "output_tokens": 2023,
  "output_tokens_details": {
   "thinking_tokens": 10
  },
  "server_tool_use": {
   "web_search_requests": 0,
   "web_fetch_requests": 0
  },
  "service_tier": "standard",
  "cache_creation": {
   "ephemeral_1h_input_tokens": 7809,
   "ephemeral_5m_input_tokens": 0
  },
  "inference_geo": "global",
  "iterations": [
   {
    "input_tokens": 2,
    "output_tokens": 429,
    "cache_read_input_tokens": 19390,
    "cache_creation_input_tokens": 565,
    "cache_creation": {
     "ephemeral_5m_input_tokens": 0,
     "ephemeral_1h_input_tokens": 565
    },
    "type": "message"
   }
  ],
  "speed": "standard",
  "fallback_credit": null
 },
 "modelUsage": {
  "claude-opus-5-5": {
   "inputTokens": 10,
   "outputTokens": 2023,
   "cacheReadInputTokens": 85707,
   "cacheCreationInputTokens": 7809,
   "webSearchRequests": 0,
   "costUSD": 0.1201134,
   "contextWindow": 1000000,
   "maxOutputTokens": 128000,
   "thinkingTokens": 10,
   "canonicalModel": "claude-opus-5-5",
   "provider": "firstParty",
   "costBasis": "list"
  }
 },
 "permission_denials": [],
 "terminal_reason": "completed",
 "fast_mode_state": "off",
 "fast_mode_disabled_reason": "sdk_opt_in_required",
 "subagent_stats": {
  "spawned": 0,
  "requested": {
   "background": 0,
   "foreground": 0,
   "unset": 0
  },
  "started_in_background": 0,
  "max_depth": 0,
  "spawned_by_subagents": 0,
  "completed": 0,
  "failed": 0,
  "killed": {
   "parent": 0,
   "user": 0,
   "system": 0
  },
  "refused": {
   "depth_limit": 0,
   "concurrency_limit": 0,
   "budget": 0
  },
  "by_type": {}
 },
 "safety_stops": 0,
 "is_error": false,
 "num_turns": 5,
 "subtype": "success",
 "api_error_status": null,
 "result": "I wrote `answers.json` and `REPORT.md`. Before computing anything I checked the CSV: it has 400 rows with unique IDs and no missing or invalid values, so nothing was dropped.\n\n**Results in `answers.json`:**\n- **`total_sessions`:** 400\n- **`total_prs`:** 256\n- **`total_api_equiv_usd`:** 4843.16\n- **`busiest_month_by_sessions`:** \"2026-09\", with 89 sessions. It only narrowly beats 2026-05, which had 88.\n- **`median_minutes`:** 88.5. With 400 sessions there's no single middle value, so this is the average of the two middle sessions.\n- **`usd_per_pr_by_tool`:** claude-code 19.5, codex 18.39. Each is that tool's total cost divided by its total PRs, not an average of per-session ratios. Sessions with zero PRs therefore cause no division problems.\n\n`REPORT.md` has the 5 bullets. They cover overall volume and cost, the monthly trend (lowest in June, then rising every month), session length, the near-even split between the two tools, and cost per PR. On cost per PR, codex is about 6% cheaper, but this doesn't account for how big or how good the PRs were.",
 "ttft_ms": 2357,
 "type": "result",
 "duration_ms": 20071,
 "uuid": "[REDACTED]",
 "ttft_stream_ms": 969,
 "time_to_request_ms": 239,
 "first_content_frame_ms": 969,
 "queued_turn_count": 0,
 "result_index": 0
}

Get new runs and ranking changes weekly