Simon Willison builds a newsletter archive for his blog by talking to Codex while cooking dinner

• Setup: the ChatGPT desktop app, Codex tab, "Start new voice chat", with the model set to GPT-6 Astra High, working on a local checkout with a dev server preview open in a browser (post). The reviewed sources don't explain what the "High" label means.
• How voice is split from coding: voice runs on GPT-Live, and the coding model is chosen separately. A July analysis said the desktop app coordinated tasks with GPT-5.6 Terra; that was written before Astra launched (Digital Applied).
• Astra API pricing: $10 per million input tokens and $50 per million output tokens, with a staged rollout (Scalevise).

• What was built: a standard Django feature: a new model, migration, admin configuration, views and templates. Newsletters were deliberately kept off tag pages and the homepage but shown on date archives. Only the unique monthly newsletters were added to search (post).
• Importers:
• Substack RSS for recent issues.
• Substack's undocumented /api/v1/archive endpoint for older ones. The agent ran a web search to work out how to paginate it and found Karen Spinner's write-up, which recommends offset paging with limit=12.
• Two GitHub repositories, one public and one private.
• Problem caught in review: the agent's importer ran git in a subprocess, which "won't work on Heroku" without extra build configuration. It was replaced with token-authenticated GitHub API calls (PR #719).
• Reproducibility: the full transcript, including the speech disfluencies, and the commit history are public (Gist).

• Claim: an agent can turn rambling, self-correcting speech into a working spec. Evidence: the transcript excerpt and full Gist (Gist). This is one person's experience on one feature.
• Claim: the feature reached production. Evidence: the merged PR and the live index page.
• Claim: desktop voice can direct several agents at once. Evidence: OpenAI's own description (OpenAI Community). Willison's session used a single agent, so it doesn't test this.
• Claim: Astra is OpenAI's "most aligned" model. Evidence: a statement from OpenAI's CEO at the launch event (live blog). No independent test confirms it.

• Willison has long used ChatGPT voice on his phone while walking the dog, mostly for research and brainstorming. He says the desktop version is different because it adds a visual preview and lets him fall back to the keyboard (post).
• He earlier documented adding a content type to his blog through agent prompting in his agentic engineering guide (guide).
• OpenAI regularly uses voice-driven demos at DevDay (live blog).

• Voice runs out at the detail stage. For pasting errors, giving examples and pointing at specific code, typing remained faster. Talking to a computer is also impractical in a shared office (post).
• The agent made a real deployment mistake. The git subprocess choice was only caught in human review (PR #719).
• Platform constraints:
• Voice usage is metered separately from Codex task budgets; on Plus that's roughly 15–30 minutes per 5-hour window.
• There's no public GPT-Live API.
• Screen-context features work only on macOS.

(Digital Applied)

• Credentials: a human still had to create the read-only token for the private repository (PR #719).

• Voice adds time, not speed. It turns otherwise idle time into specification and scaffolding work; the change is in when work can happen, not in how fast it gets done.
• Keep code review in place. In this case, review caught a problem that would have broken deployment.
• Expect a mixed workflow: voice for requirements, then typing for review and fixes.
• Plan around the limits. On lower plans the voice metering caps sessions, and without a GPT-Live API, teams can't build their own voice integrations yet (Digital Applied).
• A cheaper model option: OpenAI positions GPT-6.1 Sol as close to Astra in capability at one-fifth of Astra's token prices (search listing for OpenAI's Sol announcement).