Ask HN: Do you use Claude Code, Codex, or something else?

The Developer Tooling Divide: Claude Code vs. Codex

A recent Hacker News thread posed a deceptively simple question to the engineering community: "Do you use Claude Code, Codex, or something else?" The responses reveal a rapidly fragmenting landscape in AI-assisted development, where tool choice is no longer just about code completion but about fundamentally different philosophies of human-machine collaboration.

What Happened

The thread surfaced a growing schism in how developers integrate AI into their workflows. Claude Code (Anthropic's agentic coding tool) and Codex (GitHub's Copilot evolution) represent two competing paradigms. Users reported that Claude Code excels at autonomous, multi-step reasoning tasks—such as refactoring entire codebases or debugging complex logic chains—while Codex remains preferred for inline suggestions and real-time pair programming. Notably, several commenters mentioned abandoning both for lighter-weight alternatives like Aider or local models, citing cost, latency, or control concerns.

Why It Matters

This isn't a trivial "which tool is better" debate. It signals a maturation of the AI coding market where practitioners are making deliberate trade-offs:

• Agentic vs. Assistive: Claude Code operates as an autonomous agent that can plan, execute, and iterate without constant prompting. Codex (and Copilot) remain largely reactive, completing lines or blocks as you type. The choice reflects whether developers want a collaborator or a delegate.

• Cost calculus: Several HN users noted that Claude Code's token consumption for complex tasks can become expensive quickly, while Codex's per-seat pricing feels more predictable. This mirrors enterprise concerns about AI operational costs.

• Trust and transparency: Developers who switched to local models (e.g., CodeLlama, DeepSeek) cited data privacy and the ability to audit model behavior. This suggests that for production code, some teams prioritize explainability over raw capability.

Implications for AI Practitioners

The thread underscores three practical realities:

• No one-size-fits-all tool exists. The most effective practitioners are building layered workflows—using Claude Code for architectural reasoning, Codex for boilerplate, and traditional IDEs for debugging.

• The "vibe coding" era is ending. Early adopters who used AI indiscriminately are now developing heuristics for when to delegate versus when to write code manually. This is a healthy professionalization of AI-assisted development.

• Tool switching costs are real. Several commenters described painful migrations between tools, losing context and custom configurations. This suggests that platform lock-in may become a competitive moat for Anthropic and GitHub.

Key Takeaways

• The Claude Code vs. Codex debate reflects a fundamental choice between autonomous agentic workflows and reactive assistive tools, not just brand preference.
• Cost, latency, and data privacy concerns are driving a subset of developers toward local models and lighter alternatives.
• Mature AI practitioners are building hybrid workflows, using different tools for different cognitive loads rather than relying on a single solution.
• Tool switching friction is emerging as a hidden cost, suggesting that ecosystem stickiness will shape the next phase of the AI coding market.