SYLIN

Codex

by OpenAI · Agent · Text & code

promising

OpenAI's coding agent; early users report good results, mostly with GPT-6.2.

OpenAI reinstated the 5-hour usage limit for Codex on Plus and Business standard plans, on top of the weekly limit

Why we say so

Simon Willison used the desktop app with GPT-6.2 Astra to build a comparison video, and AI Search calls it the best harness for GPT 6.2. An HN commenter says it unlocked and reimplemented paid Android app features offline. These are positive but few, and no independent benchmarks are cited.

Evidence

  1. Nov 19
    A user had Codex unlock premium features of subscription Android apps and reimplement them offline, with no extra skills
  2. Sep 9
    OpenAI case study: Codex running GPT-5.6 Sol helps run quantum computing lab experiments
  3. Sep 8
    Verbosity reported as a smaller problem with Codex than with Claude
  4. Sep 7
    One commenter says Codex limits are much better than Claude's, and the model smarter
  5. Sep 3
    Armature's 17k-session study found Codex almost always searches the web before choosing tools
Show all 8
  1. Aug 27
    When run on AWS Bedrock, Codex reportedly gets almost no prompt-cache reads and heavy cache writes, causing roughly 10x expected charges
  2. Aug 22
    A user who downgraded Max to Pro over Fable's perceived quality drop calls Codex with GPT 5.6 Sol very good
  3. Aug 21
    Reports conflict: some say Codex over-engineers (25,000+ line diff on a small ticket, excess provenance gates); others say Claude does

Alternatives

Claude CodeIn ~17k measured sessions, Claude Code rarely searched the web, while Codex almost always did and Cursor sat in between
CursorArmature's study found Cursor searches the web moderately, between Claude Code and Codex
GLMCommenters say security researchers are shifting to Chinese models because US frontier models refuse exploit work
OpenClawCommenter says a Codex or Claude Code instance on a VPS, with bridges it writes itself, replaces OpenClaw in minutes