Skip to content

Claude Haiku 5.5 (effort default (no --effort flag) · thinking on) on Anchor drag alarm from a GPS track

AClaude Haiku 5.5default · thinking on · Claude Code 2.1.299 (headless -p) · cloud · 2026-10-20 · by @bobsiopener

Trust 80%Fully autonomoussuccess · first try
Estimated cost
$0.0081
Wall time
34s
Human step-ins
0
Outcome
Pass

Compare with other models on this task →

Summary

Exact prompt

# Anchor drag alarm from a GPS track `input.json` (synthetic, not for real anchor watch): `anchor` [lat, lon], `rode_m`, `depth_m`, `bow_height_m`, `gps_offset_m`, `margin_m`, `consecutive` N, and `track` (CSV `t,lat,lon`, t in seconds). Write `solve.py` (Python 3 standard library only); `python solve.py input.json answer.json` writes `answer.json` and `anchor.svg`. Fresh inputs are used too. ## Rules 1. Alarm radius = sqrt(rode^2 - (depth + bow_height)^2) (0 if rode <= depth + bow_height) + gps_offset + margin. 2. Track rows: skip the header. A row is **bad** (counted, skipped) if it has fewer than 3 fields, t is not an integer, lat/lon is not a finite number in range [-90, 90] / [-180, 180], or t is not greater than the last *valid* row's t. 3. Distance from the anchor (metres): R = 6371000; m = mean of the anchor and point latitudes; x = radians(dlon) * cos(m) * R (east), y = radians(dlat) * R (north); distance = hypot(x, y). A point is **outside** if distance > radius. 4. An **event** is a maximal run of consecutive valid points that are outside, with length >= N. Report `start_t` (first point), `trigger_t` (Nth point), `end_t` (last point), `max_dist_m`, and `bearing_deg` = atan2(x, y) mod 360 from the anchor to the run's last point. 5. Round the radius to 2 decimals and distances and bearings to 1. ## Visual output: anchor.svg Static, self-contained SVG (root `<svg>`, no scripts, images, links or external assets; renders headless) with `<circle id="swing-circle">`, `<polyline id="track">` and an element `id="anchor"`. The drawing itself is not graded. ## answer.json `{"radius_m", "valid_points", "bad_rows", "events": [...], "max_dist_m", "outside_points"}` ## Scoring Hidden checks (partial credit; all must pass): `swing_radius` (within 0.011), `track_cleaning` (exact), `distances`, `alarm_events` (times exact, values within 0.11), `anchor_svg_static` (anchor.svg exists, follows the static rules, has the ids/tags above), `fresh_run_1`, `fresh_run_2`. 60-second limit per run.

Catenary-free swing radius with offsets, rejecting bad and out-of-order rows against the last valid time, short excursions that must not trigger, and a real drag run with trigger time and bearing. 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).

Benchmark code for comparing AI models, not navigation or safety software.

Success criteria: 1. swing_radius 2. track_cleaning 3. distances 4. alarm_events 5. anchor_svg_static 6. fresh_run_1 7. fresh_run_2

Setup notes: Use Case Scout batch 10 case k07-anchor-drag-alarm (offline fixture). Prepared with batch10/prepare.sh; hidden key/tests kept outside the agent's work dir. Prompt = TASK.md verbatim. Scored with the scout's scorer: `python score.py WORKDIR KEYDIR`. Tools: Read/Write/Edit/Glob/Grep/Bash, web disabled. Single run, no retries. Visual output (anchor.svg) rendered offline to PNG with google-chrome headless, network blocked, 1280x800 at ~3 s after load. Claude Code's own reported cost for claude-haiku-5.5: $0.008096. Score: 7/7. 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.

Output

Metrics

Estimated cost
$0.0081
Wall time
34s
Human step-ins
0
Turns
8
Model usage
ModelInput tokOutput tokCost
AClaude Haiku 5.5anthropic/claude-haiku-5.5effort default (no --effort flag) · thinking on194,1866,378$0.0081

Reproduce this run

  1. Use the use case prompt on the use case page.
  2. Same setup: Claude Haiku 5.5 (anthropic/claude-haiku-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: 1. swing_radius 2. track_cleaning 3. distances 4. alarm_events 5. anchor_svg_static 6. fresh_run_1 7. fresh_run_2
  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": 32757,
  "stop_reason": "end_turn",
  "session_id": "[REDACTED]",
  "total_cost_usd": 0.00809573,
  "usage": {
   "input_tokens": 16,
   "cache_creation_input_tokens": 15597,
   "cache_read_input_tokens": 178573,
   "output_tokens": 6378,
   "output_tokens_details": {
    "thinking_tokens": 1853
   },
   "server_tool_use": {
    "web_search_requests": 0,
    "web_fetch_requests": 0
   },
   "service_tier": "standard",
   "cache_creation": {
    "ephemeral_1h_input_tokens": 15597,
    "ephemeral_5m_input_tokens": 0
   },
   "inference_geo": "global",
   "iterations": [
    {
     "input_tokens": 2,
     "output_tokens": 589,
     "cache_read_input_tokens": 27553,
     "cache_creation_input_tokens": 583,
     "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 583
     },
     "type": "message"
    }
   ],
   "speed": "standard",
   "fallback_credit": null
  },
  "modelUsage": {
   "claude-haiku-5-5": {
    "inputTokens": 16,
    "outputTokens": 6378,
    "cacheReadInputTokens": 178573,
    "cacheCreationInputTokens": 15597,
    "webSearchRequests": 0,
    "costUSD": 0.00809573,
    "contextWindow": 1000000,
    "maxOutputTokens": 128000,
    "thinkingTokens": 1853,
    "canonicalModel": "claude-haiku-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": 8,
  "subtype": "success",
  "api_error_status": null,
  "result": "I wrote `solve.py`, and it produces `answer.json` and `anchor.svg` for the given input in about 0.02 s. The SVG parses and contains the `swing-circle`, `track` and `anchor` ids. I have not seen the hidden checks, so the scores are unverified.\n\nResults for `input.json`:\n- **Radius:** 38.15 m (sqrt(30\u00b2 \u2212 7.1\u00b2) + 4 + 5).\n- **Track:** 290 valid points, 10 bad rows. I checked the rejected rows: 7 have empty or `NaN` lat/lon, and one (t=2718) is not greater than the previous valid t.\n- **Events:** 3 (t=822\u2013902, 2640\u20132967, 2988\u20133071). Max distance over all valid points is 50.5 m, with 49 outside points.\n\nA fresh input with different parameters and a relative track path also ran cleanly.\n\nChoices I made where the rules were ambiguous, so you can redirect me:\n- **Header:** the first line is always skipped.\n- **Blank lines:** counted as bad, since they have fewer than 3 fields.\n- **Event runs:** built over the cleaned sequence, so a bad row does not break a run.\n- **`max_dist_m` (top level):** the maximum over all valid points, not just outside points.\n- **Track path:** if the `track` path in the input doesn't exist as given, it is resolved relative to the input file's directory.\n- **`outside_points`:** a count of points with distance > radius, which I added to the answer alongside the required fields.\n\nI also regenerated `anchor.svg` in the work directory, replacing the previous one.\n\nFiles are in `./usecase-runner/runs/k07-claude-haiku/work/`:\n- solve.py\n- answer.json\n- anchor.svg",
  "ttft_ms": 1342,
  "type": "result",
  "duration_ms": 33306,
  "uuid": "[REDACTED]",
  "ttft_stream_ms": 1148,
  "time_to_request_ms": 256,
  "first_content_frame_ms": 1148,
  "queued_turn_count": 0,
  "result_index": 0
 }
]

Get new runs and ranking changes weekly