What each product actually is
Amp positions itself as a terminal-first agentic coding CLI (with related web and IDE-connected surfaces in the current product story). The review thesis is precision in large codebases: reason about the repository, then edit with the discipline of a senior engineer rather than spraying speculative patches. Buyers who already live in the shell and care about code intelligence are the core audience.
Codex is OpenAI's multi-surface coding agent. Catalog and review framing emphasize writing, review, debug, refactor, and automation across app, editor, terminal, and cloud/async tasks. Managed ChatGPT-plan access covers some surfaces; API-key usage covers CLI, SDK, and IDE paths. The loop is often "delegate a well-scoped task and review the result," not only continuous terminal pairing.
Scoreboard with dates (no re-test)
Amp (tested 2026-03-25): overall 85, speed 87, privacy 83, developer experience 84 (v0.9.4).
Codex (tested 2026-03-25): overall 80, speed 72, privacy 68, developer experience 79 (v0.9.4).
Amp leads on every published Score v1 dimension in these two reviews. The shared test calendar day (2026-03-25) is fine to state on the page; it is not evidence of a single dual session or a fresh re-run for this comparison.
Workflow fit — terminal precision vs multi-surface delegation
Choose Amp when the daily loop is terminal-native agentic editing against a large repo and you want the Score v1 evidence for speed (87) and DX (84) on that path. The Amp review frames it as the tool to beat for complex systems work in the shell.
Choose Codex when you want OpenAI's broader surface area: managed app/cloud tasks, editor integration, and async sandboxed runs for well-defined jobs. The Codex review is explicit that it trades some real-time interaction for autonomous execution — a different fit than Amp's terminal precision story, even when both sit in ai-cli-agents.
Privacy and billing posture
Amp's privacy score is 83 (tested 2026-03-25). Catalog pricing describes freemium monthly tiers (including Megawatt around $20/month with included usage and Gigawatt around $200/month) plus pay-as-you-go API-style billing. Model access and usage limits still matter; treat the published review and catalog summaries as the evidence base, not invent new plan math.
Codex's privacy score is 68 (tested 2026-03-25), reflecting OpenAI-managed and API-key paths rather than a self-hosted story. Catalog pricing ties managed surfaces to ChatGPT subscription plans and bills CLI/SDK/IDE API-key usage on token rates. Teams already standardized on ChatGPT/OpenAI will often prefer that commercial shape even when Amp wins the Score v1 table.
Who should pick which — and the winner guide
Pick Amp if you want a terminal-native agentic CLI with the stronger dated Score v1 overall (85) and you care about large-repo code intelligence. Pick Codex if OpenAI multi-surface / async cloud coding is the constraint and ChatGPT or API access is already how you buy models.