GPT-5 vs GPT-OSS-120B in finance

Every finance benchmark where both have a score (OpenAI and OpenAI). Data refreshed 6 Nov 2026.

Shared benchmarks
2
GPT-5 only
5
GPT-OSS-120B only
1

Shared benchmarks, side by side

Each benchmark is drawn on its own scale. Bars are not comparable across rows.

  • GPT-5
  • GPT-OSS-120B

  • CAIA (Crypto AI Agent Benchmark) accuracy % (majority vote, 5 runs), higher is better

    GPT-5:67.4 (better)
    GPT-OSS-120B:62.9

  • StockBench final return %, higher is better

    GPT-5:0.300 (better)
    GPT-OSS-120B:-0.900
Benchmark
GPT-5
GPT-OSS-120B

9 benchmarks, 2 scored for both. Bold marks the better score where both have one.

CAIA (Crypto AI Agent Benchmark)accuracy % (majority vote, 5 runs), higher is better
67.4
Paper Tables 2-3 (Reported by the authors, 17 Jan 2026)
62.9
Paper Tables 2-3 (Reported by the authors, 17 Jan 2026)
StockBenchfinal return %, higher is better
0.300
Paper Table 2 (Reported by the authors, 2 Mar 2026)
-0.900
Paper Table 2 (same as project leaderboard) (Reported by the authors, 6 Nov 2026)
AI-Tradercumulative return %, higher is better
1.6
Paper Table 1 (Reported by the authors, 1 Dec 2025)
Not scored
Reasoning Models Ace the CFA Exams (CFA mock-exam evaluation)Level I accuracy %, higher is better
96.1
Paper Table 5 (Reported by the authors, 9 Dec 2025)
Not scored
Reasoning Models Ace the CFA Exams (CFA mock-exam evaluation)Level III constructed-response score %, higher is better
71.8
Paper Table 5 (Reported by the authors, 9 Dec 2025)
Not scored
K-FinHallubinary detection F1, higher is better
0.823
Paper Table 3 (Reported by the authors, 28 May 2026)
Not scored
LiveTradeBenchcumulative return %, higher is better
5.3
Paper Table 2 (Reported by the authors, 5 Nov 2025)
Not scored
PolyBenchConfidence-Weighted Return (CWR) %, higher is better
Not scored
-17.8
Paper Table 5 (Reported by the authors, 3 Apr 2026)
PRBench Financeweighted rubric score (0-100), higher is better
51.3
Scale Labs leaderboard (Official board, 6 Nov 2026)
Not scored

Scores from different benchmarks use different scales and are not added up. Settings can differ between sources; see each leaderboard for details.