GPT-5.4 (2026-03-05): benchmark results
March 5, 2026 GPT-5.4 snapshot, tracked when sources report the dated API model. Provider: OpenAI. Released 2026-03-05. Access: API.
Aggregate Rating 1809 ± 8, rank #59 of 1656 rated models, from 462 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AISI Cyber CTFs (inference-scaling export) @ 50M tokens | 84.81 | Cumulative flag-capture rate within 50M total tokens (%; mea | 100 |
| Arabic Dialect SME Rubrics - Judge Deviation | 10.21 | Judge mean absolute deviation (percentage points) between th | 100 |
| EgoGapBench | 66.1 | Accuracy (%; zero-shot multiple choice on 1,000 single-image | 100 |
| EgoGapBench - Direct Actions | 50.9 | Accuracy (%; zero-shot multiple choice on 650 direct-action | 100 |
| EgoGapBench - Four Options | 65.7 | Accuracy (%; zero-shot multiple choice on 382 four-option si | 100 |
| EgoGapBench - Indirect Actions | 94.3 | Accuracy (%; zero-shot multiple choice on 350 indirect-actio | 100 |
| EgoGapBench - Three Options | 66.3 | Accuracy (%; zero-shot multiple choice on 618 three-option s | 100 |
| Horangi 4 - IFEval-Ko | 95 | Instruction-following accuracy (%) | 100 |
| ImmersedPrivacy - Social Context Rating Error (Audio as Text) | 1.03 | Mean absolute error (Tier 2: the model's 1-5 appropriateness | 100 |
| LLM-jp4 Original Release Judge - MT-BenchEN | 8.89 | Judge score | 100 |
| LLM-jp4 Original Release Judge - MT-BenchJA | 8.98 | Judge score | 100 |
| LLM-jp4 Original Release Judge - llm-jp-instructions | 4.83 | Judge score | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-5-4-2026-03-05 · How It Works · Data refreshed daily, snapshot 2026-10-20.