Codex

By OpenAI · Available on HarnessRouter

OpenAI’s coding agent for building, reviewing, testing, and shipping software across local and cloud development environments.

30-day usageAug 14–Sep 12, 2026

Data as of September 13, 2026 · About the data

Rank#9#8 of 21 CLI agents · 7-day #8

Tokens3.8T

Requests53.6M

Capabilities and verification

Reviewed against the HarnessRouter capability catalog on August 24, 2026. Every verified row cites its public source; rule-derived rows apply HarnessRouter's published classification rules to documented behavior. A “No public evidence” badge marks a gap in the public record, not a statement that Codex lacks them.

Configuration & support

Other modelsDirectly verified
Custom instructionsDirectly verified
SkillsDirectly verified
Custom toolsNo public evidence
MCPDirectly verified
Sandbox & approvalsDirectly verified
SessionsDirectly verified
SubagentsDirectly verified

Execution & agency

Workspace filesDirectly verified
Structured actionsDirectly verified

Environment control

Browser controlDirectly verified
Computer controlDirectly verified

Observation & capture

Screen observationDirectly verified
Screenshot exportDirectly verified
Recording exportNo public evidence

Artifact outputs

DocumentsDirectly verified
SpreadsheetsDirectly verified
PresentationsDirectly verified
ImagesDirectly verified
VideoRule-derived
App deploymentDirectly verified

Sources: openai.com · developers.openai.com

Machine interfaces

Headless CLI · Stable

Benchmark evidence

Codex configurations from the HarnessRouter Care Prep benchmark, published August 10, 2026. One controlled synthetic-data task; results apply to this test, not universally.