ChatGPT Work
Full referenceChatGPT Work is the product surface. GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna are selectable models within it; individual test turns are named by the model selected for that turn.
What the authenticated client exposes
| Surface | September 24, 2026 observation |
|---|---|
Work model catalog, GET /backend-api/tpp/models/ |
Seven Work records: GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, GPT-5.5, GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna. The GPT-6 slugs are gpt-6-astra-wm, gpt-6-sol-wm, and gpt-6-luna-wm. None of the records has a prompt or instruction field. |
| Conversation initialization | The observed Work session defaulted to gpt-6-luna-wm. This is a selection default, not an instruction disclosure. |
| GPT-6 Luna test turn | The client request selected gpt-6-luna-wm and sent one user message. No system, developer, base-instruction, or persistent-instruction field appeared. The visible response matched the requested exact token. |
| GPT-6 Astra test turn | The client request selected gpt-6-astra-wm and sent one user message. The same instruction fields were absent. The visible response matched the requested exact token. |
| GPT-6 Sol test turn | The client request selected gpt-6-sol-wm and sent one user message. The same instruction fields were absent. The visible response matched the requested exact token. |
| Work voice prefetch | An automatic /realtime/wm prefetch selected chat_mode: "work", backend_model: "gpt-6-luna-wm", voice_mode: "wingman", and client_tools: []. This was not a completed microphone call. |
The sanitized three-model request record is published beside this report. Request identifiers, conversation identifiers, account data, and other private browser state were excluded.
What is not recoverable at this boundary
The browser client does not transmit the Work service's system/developer instruction stack in the observed requests. That instruction layer can be assembled or injected server-side after the client request. Therefore:
- the three observations prove the selected GPT-6 model slugs and the absence of client-visible instruction fields;
- they do not prove that the models have no persistent instructions;
- they do not establish that the three Work models share one server-side prompt;
- they do not turn Codex model-catalog instructions into ChatGPT Work instructions.
The loaded Work JavaScript assets contained no You are ChatGPT base prompt. Prompt-like You are Codex strings recovered from the desktop package belong to Codex desktop helpers, app context, or voice orchestration and are published separately.
Separate installed-app and cache check
The native ChatGPT/Codex desktop package at /Applications/ChatGPT.app was inspected separately from the live Work browser session. A Safari web-app wrapper at ~/Applications/ChatGPT.app and the native app's system-hints and models cache response files were also checked without reading conversations or drafts. The system-hints cache contained routing and UI hint data; no GPT-6 Work instruction text was recovered. The model cache produced no prompt-like or codename string hit. These are negative local-artifact findings, not proof about server-side configuration.
Historical names such as Aeon, Daybreak, Mewfour, Nathree, and Persian were used only as discovery terms. They are not asserted as current Work model identities.