Codex desktop app appears to be dropping Fast mode on API keys

The Codex desktop app, when signed in with an API key, replaces service_tier=fast with service_tier=None in the requests it sends. That’s the finding of an August 13 post on Denta, whose author backs it with request logs and a throughput benchmark. On an API key, the desktop app shows no Fast toggle, and editing the Codex TOML config has no effect on it.
The same API key gets Fast in the Codex CLI. To measure the difference, the author built a tool called codex-tps and ran nine configurations, all on GPT-5.6-sol at high reasoning effort. These are the median tokens per second:
- CLI, API key, Fast: 90.2
- Desktop, API key, patched to force the Fast UI state: 87.3
- Stock desktop, API key, service tier plus feature flag: 60.7
- Stock desktop, API key, no Fast settings: 59.3
- CLI, API key, Standard: 57.5
- Stock desktop, API key,
service_tier=fast: 54.7 - Desktop, ChatGPT account, Fast/Priority: 45.6
- Stock desktop, API key, feature flag only: 45.4
- Desktop, ChatGPT account, Standard: 35.3
Every stock desktop run on an API key falls between 45.4 and 60.7, so none of them beats the CLI’s Standard run by much, whatever the settings. Only two runs pass 85, the CLI in Fast and the patched desktop app. The CLI’s Fast run comes out about 1.6 times faster than its Standard run. That roughly matches the 1.5x speedup OpenAI’s Codex documentation gives for GPT-5.6 in Fast mode.
The same documentation says to make Fast persistent by setting service_tier = "fast" and [features].fast_mode = true in config.toml. In the author’s table that is the “tier + feature flag” row. On stock desktop it scored 60.7, against 59.3 with no Fast settings at all, so the documented config makes no real difference there. At the API level, OpenAI’s API docs say Priority processing was renamed Fast mode on July 30, and a request asks for it with either "priority" or "fast" in the service_tier parameter, or through a project-level tier setting. A request that goes out with None in that field isn’t asking for either.
The author calls it a bug rather than a deliberate restriction, because Fast works with an API key in the CLI. “I’m unsure why this is happening,” the post says. The Codex docs describe Fast as available in the desktop app, the CLI and the IDE extension when you sign in with ChatGPT. On an API key, Codex bills at API token prices, and the ChatGPT credit multipliers don’t apply. The ChatGPT-account runs in the table are the slowest of the nine: 35.3 in Standard and 45.6 in Fast on the desktop app.
For now, the CLI is the supported route. There, /fast switches the mode on, and the config file keeps it on. To get the desktop app to send Fast requests, the author modified the Electron app’s code itself and published the patch on GitHub as codex-fast-mode-patcher, with a warning attached: “This is dangerous. Don’t blindly let your agent do this.”
Up next
OpenAI’s second chief revenue officer in under a year
OpenAI has a new chief revenue officer: Dali Rajic, president and COO of Wiz, replacing Denise Dresser, who lasted nine months. COO Brad Lightcap and Fidji Simo, who ran AGI deployment, have also left in the past month.
Related stories
- Codex and ChatGPT Work now share the same projects and chats
- GPT-6.1 Sol jumps from 5.4% to 86.6% on COBOL migration
- A daily benchmark checks if Codex quietly got worse
- Microsoft’s MAI Code 2 claims 96% of Opus 5.5 at 5% of price
- Claude Haiku 5.5 spotted a data leak GPT-6.1 Sol missed
- Codex helped fix a 2007 Radeon video bug for Linux 7.5
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

