// side by side · same rules, same day

Claude Code vs Codex

Claude Code 5/5 · Codex 5/5

Download the card Claude Code dossier →Codex dossier →

01

The three axes

the rules →
Claude CodeAnthropic AGENTCodexOpenAI AGENT
CAPABILITYyes / no / uncertain5/5 AGENT5/5 AGENT
PROOFhow well you can check usTRACE VERIFIEDOPERATOR INSPECTABLE
TRACE ACCESSwho can read the recordOPERATOR-OWNEDWhoever runs it holds the record; nothing is publishedOPERATOR-OWNEDWhoever runs it holds the record; nothing is published
LEASHtool calls between human turns15 median · max 9242 Claude Code sessions · one operator · median 15 · max 92not measured — measure it yourself
CHECKEDversion and datedocs as of 2026-08 (CLI 2.1.x)2026-08-11Codex CLI + cloud, docs as of 2026-082026-08-11
CONFIDENCEours, in this dossier9/108/10
COMMUNITYwhat visitors ruledA0 · W0 · C0n=0 raw votes · signal at 25A0 · W0 · C0n=0 raw votes · signal at 25
IN COURTopen challengesno cases filedno cases filed
DISPUTEDwhere the axes disagreenothing disputed — every answer is inspectablenothing disputed — every answer is inspectable
02

Question by question

rows marked "they differ" are where the comparison earns its keep
Claude CodeAnthropic AGENTCodexOpenAI AGENT
LOOPSruns until done ✓ yes TRACE VERIFIED confidence 9/10 · checked 2026-08-11 Docs' own example: 'write tests for the auth module, run them, and fix any failures' from one prompt. ✓ yes OPERATOR INSPECTABLE confidence 8/10 · checked 2026-08-11 OpenAI docs describe one agent loop that reasons, calls tools and runs tests, iterating until they pass.
CHOOSESpicks its own tools ✓ yes TRACE VERIFIED confidence 9/10 · checked 2026-08-11 Picks tools at runtime from Bash/Read/Edit plus any MCP servers; --allowedTools only bounds the set. ✓ yes OPERATOR INSPECTABLE confidence 8/10 · checked 2026-08-11 The model issues shell commands and edits at runtime; sandbox_mode only bounds what those calls may touch.
ACTSchanges the world ✓ yes TRACE VERIFIED confidence 9/10 · checked 2026-08-11 Docs: it 'edits files, runs commands'; it also stages git changes, commits and opens pull requests. ✓ yes OPERATOR INSPECTABLE confidence 8/10 · checked 2026-08-11 Default sandbox is workspace-write: it writes files and executes shell commands in the workspace.
RECOVERSfixes its own errors ✓ yes TRACE VERIFIED confidence 9/10 · checked 2026-08-11 That same documented task ends in fixing the failures it just ran; docs also show root-cause bug fixing. ✓ yes OPERATOR INSPECTABLE confidence 8/10 · checked 2026-08-11 It reads command and test output in-loop and retries; docs' guidance is that it iterates until tests pass.
UNSUPERVISEDnobody watching ✓ yes TRACE VERIFIED confidence 9/10 · checked 2026-08-11 Four of the 42 scored sessions ran to completion with zero operator interventions, and non-interactive claude -p with --allowedTools runs in CI with no approval prompts. ✓ yes OPERATOR INSPECTABLE confidence 8/10 · checked 2026-08-11 approval_policy 'never' (and --yolo) removes prompts; Codex cloud runs whole tasks in an isolated container.

Every answer here is the same one on the dossier and in the dataset — this page holds no numbers of its own. Think one is wrong? Bring a trace.