HN
Hacker News • 2일 전
GPT-6 Astra, 자동차 주행 능력 획득
IMP 8/10
핵심 요약
해커뉴스에 올라온 자율주행 리더보드에서 GPT-6 Astra(Codex·medium)가 3회 시도 중 2회째에 100% 코스 완주(5분 22초)를 달성하며 최상위를 기록했습니다. Claude Fable 5.1은 최고 45%, Grok 4.6은 11%, GPT-5.6 Sol은 6%에 그쳐 모델 간 실물 차량 제어 능력 격차가 뚜렷하게 드러났습니다.
번역된 본문
리더보드: 하나의 연속된 채팅에서 최대 3회 시도 가능. 모델을 펼쳐 각 시도를 확인하고, 시도를 클릭하면 트레이스(trace)와 영상을 볼 수 있습니다. 순위는 동일한 컨텍스트에서 최대 3회 시도 중 최고 성적 기준입니다.
【GPT-6 Astra (Codex · medium)】 최고 성적: 100% 완주, 완주 시간 5분 22초 (1회차 49%, 2회차 100% · 5:22, 3회차 —)
- 진행률(progress): 코스 중심선에서 4m 이내를 유지하며 얼마나 전진했는지를 완주 구역까지의 중심선 길이 대비 비율로 표시. 충돌 시 그 이전까지의 진행률만 인정됨.
- 거리(distance): 첫 번째 승인된 set_motion 명령부터 시도의 마지막 조작 종료까지 GPS 속도를 적분한 값.
- 완주 시간(finish time): 첫 승인 set_motion부터 마지막 조작 종료까지. 완주한 주행만 표시되며 DNF는 미완주를 의미.
- 명령(commands): 시도 중 승인된 set_motion 및 stop_now 호출 횟수.
- 토큰·비용: 사후 성찰(reflection) 포함, 정가 기준 총 토큰 및 비용.
같은 채팅 시도 내역: 1회차 49% (67.3m, DNF, 명령 8회, 120만 토큰 · $2.01), 2회차 100% (134.7m, 5분 22초, 명령 24회, 660만 토큰 · $7.74)
【Claude Fable 5.1 (Claude Code · medium)】 최고 성적: 45% (1회차 9%, 2회차 10%, 3회차 45%)
- 1회차: 9% (17.5m, DNF, 명령 3회, 58.1만 토큰 · $0.96)
- 2회차: 10% (27.3m, DNF, 명령 4회, 86.6만 토큰 · $1.35)
- 3회차: 45% (73.7m, DNF, 명령 8회, 210만 토큰 · $1.64)
【Grok 4.6 (Cursor · medium)】 최고 성적: 11% (1회차 8%, 2회차 11%, 3회차 10%)
- 1회차: 8% (14.4m, DNF, 명령 2회, 21.6만 토큰 · $0.18)
- 2회차: 11% (22.6m, DNF, 명령 3회, 29.4만 토큰 · $0.19)
- 3회차: 10% (22.2m, DNF, 명령 3회, 51.6만 토큰 · $0.29)
【GPT-5.6 Sol (Codex · medium)】 최고 성적: 6% (세 번 모두 6%)
- 1회차: 6% (15.0m, DNF, 명령 3회, 36.4만 토큰 · $0.35)
- 2회차: 6% (15.6m, DNF, 명령 4회, 72.9만 토큰 · $0.43)
- 3회차: 6% (17.1m, DNF, 명령 2회, 51.7만 토큰 · $0.27)
트랙 탐색: 궤적 재생 영상을 통해 코스를 둘러볼 수 있습니다.
원문 보기 (영어)
Leaderboard Up to 3 attempts in one continuous chat. Expand a model to inspect its runs, then click a run to view its trace and video. rank by best of up to 3 (same context) first attempt Model Best progress Finish time Attempt 1 Attempt 2 Attempt 3 GPT-6 Astra Codex · medium 100% in 5:22 49% 100% 5:22 — Expand Collapse attempt progress How far along the course centerline the attempt got while staying within 4 m of it, as a share of the centerline length to the finish zone; a collision keeps the progress reached before it. distance GPS speed integrated from the first accepted set_motion to the end of the attempt's last engagement. finish time First accepted set_motion to the end of the attempt's last engagement. Shown for completed runs only; DNF means did not finish. commands Accepted set_motion and stop_now calls during the attempt. tokens · cost Total tokens and cost at list prices for the attempt, including the reflection after it. same chat # 1 49% 67.3 m DNF 8 1.2M · $2.01 # 2 100% 134.7 m 5:22 24 6.6M · $7.74 Claude Fable 5.1 Claude Code · medium 45% 9% 10% 45% Expand Collapse attempt progress How far along the course centerline the attempt got while staying within 4 m of it, as a share of the centerline length to the finish zone; a collision keeps the progress reached before it. distance GPS speed integrated from the first accepted set_motion to the end of the attempt's last engagement. finish time First accepted set_motion to the end of the attempt's last engagement. Shown for completed runs only; DNF means did not finish. commands Accepted set_motion and stop_now calls during the attempt. tokens · cost Total tokens and cost at list prices for the attempt, including the reflection after it. same chat # 1 9% 17.5 m DNF 3 581k · $0.96 # 2 10% 27.3 m DNF 4 866k · $1.35 # 3 45% 73.7 m DNF 8 2.1M · $1.64 Grok 4.6 Cursor · medium 11% 8% 11% 10% Expand Collapse attempt progress How far along the course centerline the attempt got while staying within 4 m of it, as a share of the centerline length to the finish zone; a collision keeps the progress reached before it. distance GPS speed integrated from the first accepted set_motion to the end of the attempt's last engagement. finish time First accepted set_motion to the end of the attempt's last engagement. Shown for completed runs only; DNF means did not finish. commands Accepted set_motion and stop_now calls during the attempt. tokens · cost Total tokens and cost at list prices for the attempt, including the reflection after it. same chat # 1 8% 14.4 m DNF 2 216k · $0.18 # 2 11% 22.6 m DNF 3 294k · $0.19 # 3 10% 22.2 m DNF 3 516k · $0.29 GPT-5.6 Sol Codex · medium 6% 6% 6% 6% Expand Collapse attempt progress How far along the course centerline the attempt got while staying within 4 m of it, as a share of the centerline length to the finish zone; a collision keeps the progress reached before it. distance GPS speed integrated from the first accepted set_motion to the end of the attempt's last engagement. finish time First accepted set_motion to the end of the attempt's last engagement. Shown for completed runs only; DNF means did not finish. commands Accepted set_motion and stop_now calls during the attempt. tokens · cost Total tokens and cost at list prices for the attempt, including the reflection after it. same chat # 1 6% 15.0 m DNF 3 364k · $0.35 # 2 6% 15.6 m DNF 4 729k · $0.43 # 3 6% 17.1 m DNF 2 517k · $0.27 Explore the track View the trajectory replays ↓ Explore the course
관련 소식