Claude Code vs OpenCode

The author demonstrated that Codex creates files and runs commands without asking, while Claude Code prompts for approval for each action.

By modifying the open code JSON file, the author shows that bash commands and edits can require permission, mitigating the auto-execute default.

The author prefers Claude Code's permission prompts because they give control over every step, avoiding unwanted automatic changes.

The author highlights the animations, code preview convenience, and overall polished look as superior to Claude Code and Codex.

Demonstrated that Open Code can connect Anthropic via API, OpenAI subscriptions, Google, Open Router, and free models, unlike Claude Code and Codex which are limited to their own models.

The author criticizes this exclusivity and prefers OpenAI's approach that allows subscription usage in third-party tools.

The author considers the absence of undo a deal-breaker, as it makes recovering from mistakes difficult without manual version control.

The author shows that redo in Open Code brings back chat context but the file remains deleted, so it is not a complete Git-like redo.

The author demonstrates holding space to record voice prompts, noting it as a convenient feature for hands-free coding.

The author shows the marketplace for discovering and installing plugins and MCP servers, making Claude Code's tooling more extensible.

The author finds Codex minimal, lacking many features present in the other two agents, leading to a poor tooling experience.

Based on Reddit and Hacker News opinions, Codex's deep thinking yields thorough results, whereas Claude Code's speed suits rapid iteration.

From personal experience, Claude Code with Opus 47 yields more trustworthy code output in a single prompt.

He finds that GPT models make code style and architecture decisions that he personally does not agree with.

The author observes that even short, detailed prompts quickly exhaust the session tokens in Claude Code, while Codex allows more prompting within the same subscription period.

The author finds it unclear when tokens run out in Codex, whereas Claude Code provides clear usage statistics.

The author states that users get more usage out of their subscription with Codex due to slower token consumption.

The author observes differing architectural choices, with GPT models interpreting 'Python backend' literally, avoiding external web frameworks.

The author criticizes this closed practice, noting that Open Code is open source and community-driven, yet blocked from using subscriptions.

The author expresses a rift between ideological preference and practical necessity, favoring Claude Code for actual engineering tasks.

This hybrid approach leverages the best model while maximizing resource availability.