Use each surface for a defined job
OpenAI positions ChatGPT for reasoning and communication, ChatGPT Work for longer projects, and Codex for code, data analysis, and reproducible execution. A strong workflow keeps evidence, execution, and approval separate.
Reviewable workflow
- Frame: define the question, decision, allowed sources, exclusions, and success criteria.
- Gather: preserve source URLs, versions, dates, search strategy, and rejected evidence.
- Plan: state hypotheses, assumptions, methods, risks, and independent checks.
- Execute: use versioned code, data checksums, locked environments, tests, logs, and bounded permissions.
- Review: verify citations, calculations, statistics, plots, claims, safety, and applicability.
- Replicate: rerun from clean instructions or by an independent researcher.
- Communicate: disclose limitations, AI contribution, conflicts, and unresolved uncertainty.
ChatGPT Work can organize the long-running research record; Codex can implement and test code. Neither should be the sole keeper of authoritative data, code, or final artifacts.
Permission gates for Codex
Start with read-only repository and non-sensitive sample data. Require explicit approval for new data sources, network destinations, package installation, credentials, writes, compute spend, external messages, or irreversible actions.
Frequently asked questions
Does the academic program include API credits?
No. Current OpenAI help says it covers ChatGPT access, not OpenAI API credits.
Can Codex make an analysis reproducible automatically?
It can help create runnable code and records, but reproducibility still requires fixed inputs, environments, parameters, logs, tests, and independent rerun.
What should teams configure first?
Define source, data, repository, tool, action, compute, review, and publication boundaries before connecting production systems.
Official sources
Source check: August 5, 2026.