Skip to content

Claude Sonnet 5.5 (effort default (no --effort flag) · thinking off) on AIS collision risk: closest point of approach and alert levels

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

Trust 80%Fully autonomoussuccess · first try
Estimated cost
$0.06
Wall time
19s
Human step-ins
0
Outcome
Pass

Compare with other models on this task →

Summary

Exact prompt

# AIS collision risk: CPA, TCPA and alert levels `input.json` (synthetic; not for collision avoidance): `own` ship (`lat`, `lon`, `cog` degrees true, `sog` knots), `targets` (each `mmsi`, `lat`, `lon`, `cog`, `sog`), thresholds `cpa_danger_nm`, `cpa_caution_nm`, `tcpa_max_min`. Write `solve.py` (Python 3 standard library only); `python solve.py input.json answer.json` writes `answer.json` and `radar.svg`. Fresh inputs are used too. ## Rules 1. Skip a target (list its mmsi in `skipped`, ascending) if lat, lon, cog or sog is missing/null, sog >= 102.3 (AIS "not available"), or cog >= 360 (not available). Keep the rest in input order. 2. Local flat frame around own ship in nm: x = (lon - own lon) * 60 * cos(own lat), y = (lat - own lat) * 60. Velocity of a ship = (sog * sin(cog), sog * cos(cog)) knots. 3. Relative position r = target - own, relative velocity v = v_target - v_own. If |v|^2 < 1e-9, TCPA = 0. Otherwise TCPA (hours) = max(0, -(r . v) / |v|^2). CPA = |r + v * TCPA|. 4. `range_nm` = |r|, `bearing_deg` = atan2(x, y) mod 360 (true bearing from own ship), `tcpa_min` = TCPA * 60. 5. Level: `danger` if CPA <= cpa_danger and tcpa_min <= tcpa_max; else `caution` if CPA <= cpa_caution and tcpa_min <= tcpa_max; else `none`. 6. `alerts`: mmsi of danger then caution targets, each group sorted by tcpa_min then mmsi. Round range/CPA to 3 decimals, bearing to 1, tcpa_min to 2. ## Visual output: radar.svg Static, self-contained SVG (root `<svg>`, no scripts, images, links or external assets; renders headless) with an element `id="own-ship"` and one element `id="target-<mmsi>"` for every non-skipped target. The drawing itself is not graded. ## answer.json `{"targets": [{"mmsi", "range_nm", "bearing_deg", "cpa_nm", "tcpa_min", "level"}, ...], "skipped": [...], "alerts": [...]}` ## Scoring Hidden checks (partial credit; all must pass): `filtering`, `range_bearing`, `cpa_tcpa`, `levels_and_alerts`, `radar_svg_static` (radar.svg exists, follows the static rules, has `own-ship` and every `target-<mmsi>`), `fresh_run_1`, `fresh_run_2`. Tolerances: distances 0.002 nm, bearings 0.11, minutes 0.011 (fresh runs 0.011 for all numbers). 60-second limit per run.

AIS not-available sentinels (sog 102.3, cog 360), a stationary target, diverging targets with TCPA clamped to 0, longitude scaling, near-threshold crossing targets, and a two-key alert sort. 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. filtering 2. range_bearing 3. cpa_tcpa 4. levels_and_alerts 5. radar_svg_static 6. fresh_run_1 7. fresh_run_2

Setup notes: Use Case Scout batch 10 case k08-ais-cpa-tcpa (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 (radar.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-sonnet-5.5: $0.060703. 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
$0.06
Wall time
19s
Human step-ins
0
Turns
4
Model usage
ModelInput tokOutput tokCost
AClaude Sonnet 5.5anthropic/claude-sonnet-5.5effort default (no --effort flag) · thinking off60,8032,097$0.06

Reproduce this run

  1. Use the use case prompt on the use case page.
  2. Same setup: Claude Sonnet 5.5 (anthropic/claude-sonnet-5.5, effort default (no --effort flag) · thinking off) via Claude Code 2.1.299 (headless -p), cloud.
  3. It counts as a success only if it meets every success criterion: 1. filtering 2. range_bearing 3. cpa_tcpa 4. levels_and_alerts 5. radar_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": 18282,
  "stop_reason": "end_turn",
  "session_id": "[REDACTED]",
  "total_cost_usd": 0.060703099999999996,
  "usage": {
   "input_tokens": 6,
   "cache_creation_input_tokens": 8626,
   "cache_read_input_tokens": 52171,
   "output_tokens": 2097,
   "output_tokens_details": {
    "thinking_tokens": 0
   },
   "server_tool_use": {
    "web_search_requests": 0,
    "web_fetch_requests": 0
   },
   "service_tier": "standard",
   "cache_creation": {
    "ephemeral_1h_input_tokens": 8626,
    "ephemeral_5m_input_tokens": 0
   },
   "inference_geo": "global",
   "iterations": [
    {
     "input_tokens": 2,
     "output_tokens": 316,
     "cache_read_input_tokens": 20741,
     "cache_creation_input_tokens": 424,
     "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 424
     },
     "type": "message"
    }
   ],
   "speed": "standard",
   "fallback_credit": null
  },
  "modelUsage": {
   "claude-sonnet-5-5": {
    "inputTokens": 6,
    "outputTokens": 2097,
    "cacheReadInputTokens": 52171,
    "cacheCreationInputTokens": 8626,
    "webSearchRequests": 0,
    "costUSD": 0.060703099999999996,
    "contextWindow": 1000000,
    "maxOutputTokens": 128000,
    "thinkingTokens": 0,
    "canonicalModel": "claude-sonnet-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 wrote `solve.py`, and `python3 solve.py input.json answer.json` runs cleanly on the provided `input.json`. It writes `answer.json` and `radar.svg`.\n\nI only looked at the first few targets in `answer.json`. They show a `danger`, a `none` and a `caution` level. I haven't checked the numbers against an independent calculation or the hidden checks, and I haven't opened `radar.svg` to see how it renders.\n\n- **Command name:** the sandbox has no `python` command, only `python3`, so use `python3` when running it here.\n- **MMSI sort:** `skipped` is sorted numerically, and falls back to string order if the MMSIs mix types.\n- **Radar drawing:** `radar.svg` is a plain top-down plot. Own ship is at the centre (`own-ship`), and each non-skipped target is a coloured dot (`target-<mmsi>`) placed at its relative position, coloured by alert level.",
  "ttft_ms": 10821,
  "type": "result",
  "duration_ms": 18681,
  "uuid": "[REDACTED]",
  "ttft_stream_ms": 2508,
  "time_to_request_ms": 257,
  "first_content_frame_ms": 2509,
  "queued_turn_count": 0,
  "result_index": 0
 }
]

Get new runs and ranking changes weekly