A recent Reddit thread captures what many engineers are discovering in practice: OpenAI’s Codex, using GPT‑5 High, is landing as a pragmatic, high-velocity developer assistant, especially for code review and VS Code workflows. The original poster summed it up bluntly: “I’ve been using GPT‑5-high in Codex for a few days and I don’t miss Claude Code… The PR review feature (just mention @codex on a PR) is super easy to set up and works well.”
What people are praising
- Strong code ergonomics: Multiple users report Codex with GPT‑5 High feels snappier and better at following instructions for code tasks than alternatives they used, such as Claude Code or older Copilot models. One reader wrote, “Codex with GPT‑5 high is genuinely very good. Has replaced a large portion of my Claude usage.”
- Practical integrations: The VS Code Codex extension and @codex GitHub PR reviews are the most cited productivity wins. For solo developers working on complex projects (the OP tested it on an all-in-one gamified planner), the combination speeds up debugging, refactors, and PR triage.
- Workflow gains from better prompting: Users report they now spend more time writing specs because the model reliably executes them, a straightforward leverage effect.
Okay, let’s pause for a second… the thread is not universal praise
- Product fragmentation and naming confusion: Several commenters pointed out that “Codex CLI” and the Codex web extension are effectively different experiences. One high-value comment called out that OpenAI uses the same name for multiple distinct products. That matters for expectations, since the web UI is barebones while the CLI and VS Code integration behave differently.
- Quotas and session limits: Multiple users hit confusing usage limits on the $20 plan; some reported being blocked after aggressive early sessions. Several recommend Pro for heavy continuous usage, while others caution to be parsimonious and use lower-cost modes for simple tasks.
- Project awareness and scale: For large codebases with opaque imports or framework magic such as Nuxt auto-imports or Django string behavior, Codex can miss critical connections. Directory reads help, but manually grounding with key files or docs remains necessary.
- UX and extension bugs: A few users reported the extension refusing to make edits after a permissions flow, lack of branch selection for edits, or UI layout problems.
High-value comments to note
“Codex with GPT‑5 high is genuinely very good. Has replaced a large portion of my Claude usage.”
“The CLI and web agent seem to be completely different products which share very little code, OpenAI has a problem with reusing a single name for different products.”
There are also multiple reports of unexpected hourly and weekly limits and sessions being cut off, practical friction affecting adoption.
Context and what this means for teams
This is an instance of tooling catching up to everyday developer needs: integrations (IDE plus GitHub plus PR review bots) matter more than raw model benchmarks for most engineers. However, leaderboard comparisons referenced in the thread are noisy, since many omitted strong competitors like Claude Code or Gemini CLI or conflated multi-model frontends with single models. Treat those rankings with skepticism.
Actionable tips
- Try the VS Code extension and @codex for PR reviews first; it offers the highest immediate ROI.
- Ground the model: supply key docs or paste critical files, or add a web/context tool when working on framework-heavy projects.
- Watch quotas: test your expected session patterns on your subscription level and consider using medium or low modes for simple tasks to conserve quota.
- Start PR edits on a separate branch and run the assistant on smaller subsets of files to avoid missed context.
- Expect UX rough edges, and have fallback workflows such as a local CLI or alternative agents until integrations mature.
My takeaways
- Codex with GPT‑5 High is a tangible productivity tool for developers, not just a research demo. It shines in PR review and interactive coding within IDEs.
- Practical adoption hinges on integrations, quota transparency, and robustness in large projects, areas where users still report friction.
- Do not treat leaderboard claims as gospel; compare tools by the full workflow, including IDE support, repo awareness, cost, and limits, not only per-task scores.
Remember that there is often another side to the story: some users still prefer Claude for long structured writing, and other agents such as Cursor or Jules perform strongly in different contexts. Test with your codebase and CI patterns before making a hard migration.
Source
Source thread: https://reddit.com/r/OpenAI/comments/1n5iqqj/openai_nailed_it_with_codex_for_devs/
Recommendation
If you are evaluating AI copilots for engineering teams, prioritize integrations and quota behavior over benchmark headlines. The model quality is impressive, but the product experience will determine whether it replaces your current tools.
Leave a Reply