Verdict
Built from the measurements below
- GPT-5.3-Codex is more accurate: 82.5% against 51.2% on the same 80 questions, a 31.3-point gap that is statistically significant.
- gpt-oss-20b is 157× cheaper per question as billed ($0.000066 vs $0.01), reasoning tokens included.
- Output speed is similar (95 vs 103 tokens per second).
- gpt-oss-20b starts answering sooner: 0.7 s to the first token, against 1.4 s for GPT-5.3-Codex.
- GPT-5.3-Codex accepts more context: 400k tokens against 131k.
Which should you choose?
Choose GPT-5.3-Codex for the most accurate answers of the two and a 400k-token context window. Choose gpt-oss-20b for the lower cost per correct answer ($0.00013 against $0.013), a quicker first token, open weights you can run yourself and more hosts to choose from on OpenRouter (10 against 2). The widest subject gap is in engineering: GPT-5.3-Codex answered 11 of 15 correctly, gpt-oss-20b 4 of 15. With so few questions per subject, treat this as a lead to test on your own prompts, not a result.
Accuracy gap +31.3 points (z = 4.20). We call a winner only when z is at least 1.96, the usual 95% threshold. Speed, latency and price need a 10% margin.