1 firsthand and 1 secondhand accounts in the October guide. The sample is selected and not representative.
Firsthand account · Simon Willison · 2026-10-09
A Django feature shipped after a second round of review
Codex voice mode · GPT-6 Astra High
Willison reports building newsletter models, migrations, imports, archive pages and search integration with Codex voice mode, using GPT-6 Astra High. He then reviewed the pull request, changed a subprocess-based import to an API-based import, and refined the display before deploying.
What this does not establish: One experienced developer and one Django feature. He describes roughly half an hour by voice plus further typed work; this is not a timing benchmark or a general result for every Codex model.
An older destructive-action incident shows why agent permissions matter
Codex with full-access execution · GPT-5.6
Willison collected an OpenAI engineering statement describing reports of unintended file deletion involving GPT-5.6, full-access execution and temporary-directory handling. It illustrates a failure mode worth isolating when an agent can modify your files.
What this does not establish: A July report about an older model, relaying an engineering statement. It does not establish a current failure rate, an unresolved current bug, or that another provider is immune.
Listed currency: USD · Monthly billing. Region is unverified unless specified per offer. Annual discounts are excluded. These are published amounts, not a local checkout quote; confirm region, taxes and availability with the provider. Subscription allowance and API token billing are different quantities.
These are selected, attributed accounts with the author’s stated tool configuration and limits. They do not establish a general success rate or prove that a current version behaves the same way.