Repository work rather than a demo prompt
Compare a complete engineering flow: understand the repository, plan the change, edit files, run tests and deliver a reviewable diff. Include skills, MCP, local execution and cloud tasks only where they are actually intended for the target operation.
Boundary: agent permissions are production permissions
Record file paths, shell commands, network access, secrets, branch rules and approvals separately for both systems. A better diff does not justify broader permissions.
Pilot: three issues with a known target
- Use one bug fix, one small feature and one test improvement
- Start each system from the same commit and repository rules
- Measure test results, unwanted changes, review time and interruptions
- Do not combine local and cloud results in one metric
Claude side of the comparison: September 2026
A current evaluation should consider Opus 5.5 and the announced coordinated Projects. Both must actually be available in the account used; the Projects beta is not a generally available team feature. Compare the same assignment with recorded models, interfaces and permissions. The relevant outcome is an understandable, tested change with manageable review effort. A difference between two models should not be presented uncritically as a difference between two complete products.
Decision matrix
| Criterion | Claude Code | Codex |
|---|---|---|
| Agent path | Check skills, MCP and repository instructions in the target flow | Check local and cloud tasks with identical acceptance tests |
| Permissions | Log file, shell and network approvals visibly | Log sandbox, approvals and remote environment visibly |
| Acceptance | Diff, tests and manual correction time | Diff, tests and manual correction time |
Keep it verifiable
Primary sources
- Claude SonnetSource checked:
- Claude EnterpriseSource checked:
- Anthropic: commercial data and trainingSource checked:
- Anthropic: Remote MCP connectorsSource checked:
- OpenAI: ChatGPT WorkSource checked:
- OpenAI: Codex CloudSource checked:
- OpenAI: sandbox and approvalsSource checked:
- OpenAI: enterprise privacy and controlsSource checked:
- Anthropic: Claude Opus 5.5Source checked:
- Anthropic: Projects redesignedSource checked:
