Skip to content

Claude Opus 5.5 (effort default (no --effort flag) · thinking on) on QA a small app against its spec before launch and report only real bugs, where an empty list is a valid answer

AClaude Opus 5.5default · thinking on · Claude Code 2.1.299 (headless -p) · cloud · 2026-10-25 · by @bobsiopener

Trust 80%Fully autonomoussuccess · first try
Estimated cost
$0.16
Wall time
36s
Human step-ins
0
Outcome
Pass

Compare with other models on this task →

Summary

Exact prompt

# QA a small app against its spec before launch and report only real bugs, where an empty list is a valid answer `app/shop.py` is a small synthetic storefront module, and `SPEC.md` is its specification with numbered clauses (S1, S2, ...). The beta tester's notes are in `beta-feedback.txt`. Everything is synthetic and offline. Find where the code breaks the spec. You may read, run and test the code, but do not change anything under `app/`. Rules: - A bug is a call whose result (a return value, or the exception type raised) differs from what SPEC.md requires. Behaviour the spec requires is never a bug, even if it looks odd or a note calls it one; anything the spec doesn't cover is not a bug either. - Report each bug once under the one clause it breaks, with a call that shows it: the function name and a JSON list of positional arguments. Your call must actually give the wrong result when run against app/shop.py as shipped. - Report nothing you haven't confirmed. False reports count against you, and if the code meets the spec, the right answer is an empty list. - Write `out/findings.json`: a JSON list (possibly empty) of `{"clause": "S<n>", "call": "<function name>", "args": [<JSON arguments>], "expected": <what SPEC.md requires>, "actual": <what the code returns, or {"raises": "<ExceptionType>"}>}`. ## How this run works - Work on your own: nobody will answer questions in this task, and writing to `ask/` counts against the run. - Write only under `out/`. Create `out/` if it is missing. Inputs are synthetic; don't fetch anything from the network. - Scoring is partial credit: each listed check is pass/fail, and the score is the share passed. Numbers are compared to within 0.005 unless the task says otherwise.

Check a small synthetic storefront module against a numbered spec and report only confirmed deviations, each with a call that reproduces it. Each seed plants zero to three bugs; a quarter of the seeds plant none, and then an empty list is the right answer. A beta tester's notes flag behaviour the spec actually requires. Every reported call is replayed on the box against the shipped code and a spec-correct copy, so false positives and made-up evidence cost points. Original benchmark task written for usecaserank.com (not based on a specific public post). Headless, offline run on a reproducible fixture. Each run's notes say how its cost was measured (estimates are marked).

Success criteria: right_verdict found_real_bugs no_false_clauses evidence_reproduces

Setup notes: Batch 37 case ty02-prelaunch-qa-real-bugs-only (offline, synthetic inputs, seed 323015388). Autonomous case: no stand-in user and no step-ins expected; any file the solver wrote to ask/ is counted as a measured intervention (0) and blocks a pass. Prompt = TASK.md verbatim. Hidden key kept outside the solver's work dir. Scored with the case's scorer (python3 score.py WORKDIR KEYDIR, absolute paths); pass = every check true and no ask/ files. Tools: Read/Write/Edit/Glob/Grep/Bash, web disabled. Single run, no retries. 1-min load average at scoring 0.36 (8 cores). Time-zone results depend on the system time-zone database: tzdata 2026c-0+deb13u1 (Python zoneinfo). Scored only inside a bwrap sandbox (bubblewrap 0.12.0): no network, read-only system, case dir and key read-only, only a throwaway copy of the workdir writable. ty02 precision n/a (nothing reported) (no_false_clauses True), recall n/a (zero-bug seed) (found_real_bugs True); 0 clause(s) reported, 0 planted. Zero-bug seed: the empty list is correct; zero-bug seeds are excluded from the baseline max. Claude Code's own reported cost for claude-opus-5.5: $0.159115. Score: 4/4. Cost is an ESTIMATE: measured tokens x pricing.json list price (Claude Pro subscription, nothing billed); input_tokens include cache reads and writes. Single run.

Estimated cost List-price estimate; not actually billed.

Metrics

Estimated cost
$0.16
Wall time
36s
Human step-ins
0
Turns
4
Model usage
ModelInput tokOutput tokCost
AClaude Opus 5.5anthropic/claude-opus-5.5effort default (no --effort flag) · thinking on83,2672,924$0.16

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, effort default (no --effort flag) · thinking on) via Claude Code 2.1.299 (headless -p), cloud.
  3. It counts as a success only if it meets every success criterion: right_verdict found_real_bugs no_false_clauses evidence_reproduces
  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": 35184,
  "stop_reason": "end_turn",
  "session_id": "[REDACTED]",
  "total_cost_usd": 0.1591152,
  "usage": {
   "input_tokens": 8,
   "cache_creation_input_tokens": 10763,
   "cache_read_input_tokens": 72496,
   "output_tokens": 2924,
   "output_tokens_details": {
    "thinking_tokens": 1512
   },
   "server_tool_use": {
    "web_search_requests": 0,
    "web_fetch_requests": 0
   },
   "service_tier": "standard",
   "cache_creation": {
    "ephemeral_1h_input_tokens": 10763,
    "ephemeral_5m_input_tokens": 0
   },
   "inference_geo": "global",
   "iterations": [
    {
     "input_tokens": 2,
     "output_tokens": 778,
     "cache_read_input_tokens": 21892,
     "cache_creation_input_tokens": 1017,
     "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 1017
     },
     "type": "message"
    }
   ],
   "speed": "standard",
   "fallback_credit": null
  },
  "modelUsage": {
   "claude-opus-5-5": {
    "inputTokens": 8,
    "outputTokens": 2924,
    "cacheReadInputTokens": 72496,
    "cacheCreationInputTokens": 10763,
    "webSearchRequests": 0,
    "costUSD": 0.1591152,
    "contextWindow": 1000000,
    "maxOutputTokens": 128000,
    "thinkingTokens": 1512,
    "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": 4,
  "subtype": "success",
  "api_error_status": null,
  "result": "I found no bugs, so `out/findings.json` is an empty list: `[]`. I checked all 18 clauses against `app/shop.py` by reading the code and running sample and edge-case calls, and every result matched `SPEC.md`.\n\n**The three beta-tester notes all describe behaviour the spec requires:**\n- **S14:** `validate_email(\" [REDACTED:email] \")` returning valid is correct, because the spec says surrounding whitespace is stripped first.\n- **S11:** `paginate([1, 2, 3], 0)` raising `ValueError` is correct, because a page below 1 must raise.\n- **S6:** `slugify(\"Caf\u00e9 Menu\")` giving `\"caf-menu\"` is correct, because accented letters count as non-a-z characters and become a dash. The TODO comment in the code about this is a red herring.\n\n**Edge cases that also passed:**\n- **Cart totals:** the 10% coupon discount rounds half-up (50.05 gives 5.01). The 50.00 coupon and 65.00 free-shipping thresholds both behave as specified, and an empty cart works.\n- **Quantities:** they must be whole numbers from 1 to 10 or the call raises `ValueError`.\n- **Pagination:** `per_page` is clamped to 1\u201350, and a page past the end returns an empty list.\n- **Durations:** the three accepted formats return the right minutes; an empty string or minutes of 60 or more raise `ValueError`.\n- **Usernames:** the 3\u201316 length limits, starting with a letter, and rejecting reserved names in any letter case all work.\n\n**One borderline case I chose not to report:** `slugify(\"K\")`, where the character is the Kelvin sign (U+212A), returns `\"k\"`. Python's lowercasing turns it into an ASCII \"k\". The spec is ambiguous here: S6 says non-ASCII characters should become dashes, which would give `\"\"`. But S5 says letters are lowercased *first*, and that step already makes the character an ASCII \"k\". Under that reading `\"k\"` is correct. Since false reports count against the run, I left it out. If you read S5 as covering only A-Z, add it as an S6 finding.\n\nI didn't change anything under `app/`; the only file I wrote is `out/findings.json`.",
  "ttft_ms": 3023,
  "type": "result",
  "duration_ms": 35480,
  "uuid": "[REDACTED]",
  "ttft_stream_ms": 2432,
  "time_to_request_ms": 154,
  "first_content_frame_ms": 2432,
  "queued_turn_count": 0,
  "result_index": 0
 }
]

Get new runs and ranking changes weekly