Proposal
Add a separate OpenAI Codex provider for models served through ChatGPT/Codex OAuth.
The main openai provider should continue describing direct API access. The Codex provider can inherit shared model metadata with base_model and override the limits that differ on the Codex backend. Harnesses such as OpenCode can then select the correct provider when Codex OAuth is active.
This has moved beyond a one-time client workaround: four current Codex models now have different limits from their direct API definitions.
Evidence
| Models |
Direct OpenAI API |
Codex OAuth |
gpt-5.5 |
1,050K context / 922K input / 128K output |
400K context / 272K input / 128K output |
gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna |
1,050K context / 922K input / 128K output |
500K context / 372K input / 128K output |
The Codex values can be independently checked with codex debug models. GPT-5.5 reports 272000; all three GPT-5.6 models report 372000.
Related
Proposal
Add a separate OpenAI Codex provider for models served through ChatGPT/Codex OAuth.
The main
openaiprovider should continue describing direct API access. The Codex provider can inherit shared model metadata withbase_modeland override the limits that differ on the Codex backend. Harnesses such as OpenCode can then select the correct provider when Codex OAuth is active.This has moved beyond a one-time client workaround: four current Codex models now have different limits from their direct API definitions.
Evidence
gpt-5.5gpt-5.6-sol,gpt-5.6-terra,gpt-5.6-lunaThe Codex values can be independently checked with
codex debug models. GPT-5.5 reports272000; all three GPT-5.6 models report372000.Related