GPT-5.3 Codex: benchmark results
Specialized for coding and software-agent work in a terminal. Supports context compaction to carry task state through coding sessions that exceed one context window. Provider: OpenAI. Released 2026-02-05. Access: API.
Aggregate Rating 1797 ± 5, rank #69 of 1656 rated models, from 393 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AISI Cyber CTF | 75 | Success Rate (%) | 100 |
| AISI Cyber CTFs (inference-scaling export) @ 500K tokens | 40.75 | Cumulative flag-capture rate within 500K total tokens (%; me | 100 |
| BinaryAudit | 94.12 | Avg Success Rate (%) | 100 |
| Boiling the Frog - Safe Agency Score | 68.5 | Safe Agency Score (SAS, %): benign strict success rate times | 100 |
| CP-SynC-XL - MiniZinc + OR-Tools (Heuristic Prompt) | 54.1 | Correct instances (%; share of the 4,577 CP-SynC-XL instance | 100 |
| CP-SynC-XL - Python | 63.4 | Correct instances (%; share of the 4,577 CP-SynC-XL instance | 100 |
| ClawForge - Partial Credit | 82.38 | Partial-credit score (0-100): fraction of a task's evaluatio | 100 |
| ClawForge - Workflow Completion | 42.5 | Strict full-pass accuracy (%) on the 80 tasks whose primary | 100 |
| DeployBench | 51 | Task success rate (%) on all 51 research-artifact deployment | 100 |
| DeployBench - AI and ML | 52 | Task success rate (%) on the 25 AI and machine-learning arti | 100 |
| EVMbench | 51 | Unweighted mean of Detect/Patch/Exploit task means (%); no h | 100 |
| HiL-Bench Original Paper - SQL - Question Precision | 55.3 | Question Precision (%) | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-5-3-codex · How It Works · Data refreshed daily, snapshot 2026-10-20.