benchmarks
GPT-5.3-Codex is OpenAI’s coding-focused agentic model for software engineering, terminal tasks and computer-based work. OpenAI reports improved speed and stronger terminal and computer-use performance over GPT-5.2-Codex, although autonomous workflows still require testing, review and carefully scoped permissions.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| SWE-bench VerifiedResolving real GitHub issues | 74.8%25 Feb 2026 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 78.4%5 Feb 2026 | View |
Weighted mean of 2 verified benchmark results: SWE-bench Verified 74.8%, Terminal-Bench 78.4%.
Capabilities
- Handles extended software-engineering workflows involving repository exploration, implementation, debugging and testing.
- Uses terminal tools to carry out development and system tasks.
- Supports computer interaction for tasks extending beyond code generation.
- Combines coding capabilities with reasoning and professional knowledge.
- Supports interactive collaboration through progress updates and user steering during ongoing work.
What improved
- Combines GPT-5.2-Codex’s coding strengths with GPT-5.2’s reasoning and professional knowledge.
- OpenAI reports 25% faster performance for Codex users than its predecessor.
- Improves Terminal-Bench 2.0 and OSWorld performance over GPT-5.2-Codex.
- Provides more frequent progress updates and better support for steering ongoing work.
Known flaws & limitations
- Software-engineering evaluations show incomplete task success; generated changes still require testing and review.
- Computer-use performance remains imperfect, limiting dependable unattended operation.
- Agentic workflows remain susceptible to malicious instructions embedded in external content.
- Cybersecurity safeguards can restrict requests involving sensitive or dual-use security work.
Risks, safety & security issues
- Advanced cybersecurity capabilities could assist harmful activity; OpenAI applies safeguards associated with its High cybersecurity capability classification.
- Malicious instructions in repositories, documents or tool output can attempt to redirect an agent away from its authorized task.
- Broad tool permissions can amplify the consequences of incorrect or maliciously redirected actions.
- Overreliance on generated code can allow correctness or security defects to reach production.
Good fit for
- Repository-level feature development, refactoring and bug fixing with human review.
- Terminal-based development workflows with automated tests and scoped permissions.
- Long-running engineering tasks that benefit from progress updates and interactive guidance.
- Supervised computer-based tasks combining software development and professional knowledge.
Poor fit for
- Unattended production changes without tests, approval gates or rollback mechanisms.
- Security-critical code deployment without independent correctness and security review.
- Agents processing untrusted content while holding unrestricted credentials or destructive permissions.
References
Profile last reviewed 10/25/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.