A week of using Codex more than Claude

Original Article ↗ Hacker News Discussion ↗

Coders comparing OpenAI’s Codex tooling to Anthropic’s Claude Code report that Codex feels faster, more focused, and less verbose, while Claude often produces over-engineered solutions and floods code with explanatory comments that many see as noise. Others counter that Claude is better at inferring intent and providing architectural help, highlighting how much results vary by task, harness, and user skill. The broader thread surveys a fast-moving ecosystem of coding models (Sol, Fable, Grok, Gemini, DeepSeek, Kimi, etc.), with many developers mixing models and orchestration frameworks to balance cost, speed, reliability, and guardrails.

Overall Sentiment: Codex vs Claude for Coding

  • Many report shifting from Claude to Codex (GPT 5.6 Sol/Luna/Terra) for coding.
  • Main drivers: Codex is perceived as faster, more concise, more “business-like,” and less mentally exhausting.
  • Claude’s current behavior is often described as verbose, over-explanatory, and sometimes “attitude-filled,” though some still prefer its ability to infer intent and handle ambiguous tasks.

Code Comments and Verbosity

  • Strong dislike for Claude’s auto-generated comments: overly long, AI-styled prose, diff-history in comments, and “notes-to-self” that describe what code doesn’t do or past iterations.
  • Many argue these are noise, should live in PRs/commit messages, and confuse both humans and future model runs.
  • A minority argues such comments serve as “external memory” for LLMs and help future agents avoid mistakes.
  • Broad consensus: fewer AI-generated comments in codebases is better, even if comments in general can be valuable.

Overengineering vs Pragmatism

  • Several say Codex tends to overengineer: elaborate architectures, enums, validation frameworks, extreme defensiveness, and large diff sizes.
  • Others find Claude more likely to overcomplicate or “do too much,” while Codex stops closer to the requested scope.
  • Experiences are sharply contradictory; behavior seems sensitive to model variant, effort level, harness, and task type.

Models vs Harnesses vs Process

  • Repeated emphasis that behavior depends heavily on the harness (Codex CLI/TUI, Claude Code, OMP, prime-agent, agy, Hermes Agent, etc.).
  • Some advocate separating “process” (plan/implement/review loops) into deterministic runtimes, treating models as interchangeable roles.
  • Multi-agent patterns are popular: one model plans, another codes; or Claude and Codex critique each other’s work.

Quotas, Cost, and Usage Patterns

  • Mixed reports on which service is more generous. Some burn through Claude quickly; others hit Codex limits in days, especially using Sol Ultra.
  • Cheaper models (Luna xhigh, DS Flash, local models) are used as subagents or for bulk execution.
  • Many tailor model choice: Sol for routine/back-end, Opus/Fable for design or architecture, other models for speed or cost.

Other Models and Ethics

  • Grok, Gemini 3.7, DeepSeek, Kimi, GLM, Qwen, and local models are all mentioned as viable options.
  • Strong side-thread on ethics and politics of using particular providers (especially Grok), data privacy incidents, and whether avoiding certain vendors is meaningful or merely symbolic.