OpenAI Codex ‘Mega Update’ Sparks Mixed Reactions: Better Instruction-Following and Web/Git UX, But Limits Persist

Quick synopsis

OpenAI’s recent Codex refresh (the ‘mega update’ users are discussing) appears focused on quality-of-life improvements: stronger instruction-following (GPT-5–class behavior), a more polished web experience, improved Git-native workflows, and tighter integration with developer flows (GitHub tasks, per-issue fixes). Redditors reacted with a mix of enthusiasm and pragmatic concerns — many praising raw code quality, others flagging CLI ergonomics, session management gaps, Windows/VSCode sign-in quirks, and usage-limit headaches.

What people are saying (themes from the thread)

• Stronger instruction following: Several high-score comments note that GPT-5’s ability to “listen” is a major win for delegating coding tasks reliably. “GPT-5 is good at listening to instructions, which is probably the biggest key to delegating tasks to an LLM,” one top comment observed.
• Web experience wins: The web UI and Git-integrated workflow are repeatedly praised — users like launching tasks from GitHub and the per-issue fix flow.
• CLI & IDE pain points: Multiple users report the CLI/UI feels less ergonomic than competitors, lacks session management, and is harder to read. Windows users complain about terminal behavior and sign-in redirects in VSCode. JetBrains support appears absent and many ask for it.
• Model-tier differences: A recurring observation: users report gpt5-medium vs gpt5-high feel noticeably different for coding tasks. One user called gpt5-high “SOTA” and suggested medium behaves like a constrained flavor that sometimes fails to ingest context.
• Competitive positioning: People are actively comparing Codex to Claude Code, Gemini, Cursor, and GitHub/CoPilot. Some switch between tools, often using Codex for generation and Claude for review (or vice versa).
• Billing & limits: Users want clarity on which experiences consume web allotments, CLI limits, and team plan caps. Some say team limits make Codex impractical for heavy work.

Highlighted comments

1. AllezLesPrimrose: “GPT-5 is good at listening to instructions…Claude is still better with tools but with these improvements it’s become a very close-run thing between the big two and which better for coding.” This captures the overall sentiment: parity is close, but strengths differ.
2. dhamaniasad: A practical user story: switching between Claude and Codex, finding Codex more precise at catching mistakes despite a worse UI, and lamenting the lack of session management.
3. redditisunproductive: Flags an operational detail: gpt5-medium vs gpt5-high feel qualitatively different; gpt5-high is far better at following instructions and scaling with more ‘thinking’ tokens.

Context and why it matters

This thread is early product research from active developers. When a model improves instruction-following, teams change how they design prompts, test automation, and integrate LLMs into CI/CD. The fact that users are combining models (generate in one, review in another) is now common practice and signals that no single offering is a universal replacement for code workflows yet.

Practical implications for builders

• Try the web flow first — many find it the most frictionless entry point.
• For heavy, context-rich coding tasks, prefer gpt5-high where available; test gpt5-medium vs gpt5-high on your repo to validate behavior and costs.
• Use a multi-model review loop: generate with one model and validate with another (e.g., Codex and Claude Code) to catch both false positives and alarmist regressions.
• Verify session management and limits: if your workflow needs long-lived sessions or custom containers, confirm support before committing.

My takeaways

1. Codex’s update matters because better instruction-following materially improves developer delegation and automation reliability.
2. The web/Git workflow is a clear UX win — it lowers the bar for adoption versus pure-CLI experiences.
3. Operational concerns (session persistence, platform-specific sign-in issues, rate/usage caps) remain the main inhibitors to replacing incumbent tools.
4. Expect hybrid workflows: generate in one model, review in another; test model “flavors” under your own workloads.

Final note

User anecdotes vary by OS, IDE, and subscription tier. If you’re evaluating Codex for production, run a short POC focused on your repo, heavy-context prompts, and CI integration — those reveal real trade-offs faster than any headline.