Quartz 5

Codex

Properties1
tagstech, software

3 min read

Entity: Codex

Codex (often utilized via ChatGPT Work or the specialized Codex Ultra setting) is a highly popular, terminal-friendly, and task-focused artificial intelligence model and execution harness developed by OpenAI.

Core Characteristics

  • High Dependability & Speed: Codex is characterized by rapid execution, high reliability, and a seamless interface. It runs open-ended tasks with zero friction and is highly favored as an everyday driver for developers and automation engineers.
  • The Codex Harness: Unlike Anthropic’s interfaces, which often require multiple permission dialogues and security popups, the Codex harness allows agents to run to completion in a single, uninterrupted run.
  • Codex Ultra: A specialized, high-token-burn setting within the OpenAI interface that enables extensive, multi-billion-token agentic operations. In operational audits across 143 daily Codex threads, Nate B. Jones recorded 3.77 billion tokens in a single day, with 3.59 billion tokens (96%) consisting of reused-input-compounding.
  • Token Efficiency Optimization: Due to heavy context compounding in Codex threads, operators deploy skills like token-saver-skill to pre-search files, enforce answer bounds, and prevent unnecessary token consumption.
  • Rapid Adoption: In mid-2026, ChatGPT Work and Codex were gaining approximately one million users per day, surpassing the active user base of terminal-based tools like claude-code.

Operational Limitations & The Bounded Problem Trap

While Codex is highly efficient at execution, it exhibits a distinct behavioral profile when given open-ended problem-solving tasks:

  • Bounded Problem Nose: When given a completely free hand to audit workflows and select a problem to automate, Codex tends to keep itself bounded. It gravitates toward conservative, “boring,” and pre-voiced problems (such as optimizing handoff packages) rather than identifying high-leverage strategic opportunities.
  • Contrast with Fable 5: Codex lacks the “big model smell” and strategic reasoning of fable-5. While fable-5 is a hassle to operate, it excels at scouting high-value, ambiguous problems. Consequently, a common cost-saving arbitrage strategy is to use fable-5 to scout the problem and design the architecture, then execute the build using the cheaper, faster Codex or ringer systems.

Multi-Session Limits & OpenAI Symphony

  • Context-Switching Bottleneck: When running long multi-hour execution tasks in terminal environments, human operators reach a cognitive boundary managing 3 to 5 simultaneous Codex sessions before losing track of session state.
  • OpenAI Symphony Integration: To scale beyond individual terminal sessions, OpenAI deployed openai-symphony—a project board interface where Codex agents pull tickets directly from a central backlog, driving a 500% surge in landed pull requests (see progressive-context-shaping).

References

Graph View

Backlinks