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

A developer benchmarking the Codex desktop app on an API key found that the client strips service_tier=fast out of outgoing requests and sends service_tier=None, with stock desktop measuring a median 54.7 tokens per second against 87.3 on a patched build, both on GPT-5.6-sol at high reasoning effort.

The measurements were published by Denta, which ran a matrix of nine configurations across the desktop client and the Codex CLI, on an API key and on a ChatGPT account. Every run used GPT-5.6-sol at high reasoning effort, and each result is reported as a median.

At a glance

• There is no Fast mode toggle in the desktop interface when an API key is in use, and edits to the Codex TOML configuration file do not reach the application at all.
• Across the nine configurations measured with GPT-5.6-sol at high reasoning effort, medians ran from 35.3 tokens per second on a ChatGPT account in standard mode to 90.2 on the CLI in Fast.
• The same API key reaches Fast mode through the Codex CLI, so the author reads the desktop behaviour as a bug and published an Electron patch as codex-fast-mode-patcher on GitHub.

Why it matters Throughput is one of the few properties of an agentic coding client that a developer can observe directly, and the gap between the desktop and CLI numbers in this matrix is large enough to be felt on long runs. The divergence appears to be client-side rather than an account restriction, since the same key reaches Fast through the CLI, though a single benchmark does not establish that. The remaining fix on record involves editing a shipped client, which the author flags as dangerous.

A patched desktop build returned 87.3 tokens per second against 54.7 on stock

Stock desktop with an API key and service_tier=fast returned a median 54.7 tokens per second on GPT-5.6-sol at high reasoning effort. A desktop build patched by the author, with the Fast UI state forced on, returned 87.3 under the same model and reasoning setting.

Three further stock desktop runs with an API key sat between 45.4 and 60.7 tokens per second: the service tier plus a feature flag gave 60.7, no Fast settings at all gave 59.3, and the feature flag alone gave 45.4, all on GPT-5.6-sol at high reasoning effort.

The benchmark runs through codex-tps, a tool the author built for this test, and every row of the matrix is published alongside its own branch and log so that individual runs can be inspected. The CLI and ChatGPT-account figures come from the same set of measurements.

The Codex CLI reached 90.2 tokens per second in Fast mode with the same API key

Run through the Codex CLI with an API key, GPT-5.6-sol at high reasoning effort returned a median 57.5 tokens per second in standard mode and 90.2 in Fast mode. Fast mode is therefore reachable with an API key, though not from the desktop client.

The two desktop runs on a ChatGPT account produced 35.3 tokens per second in standard mode and 45.6 with Fast/Priority selected, both on GPT-5.6-sol at high reasoning effort. Those are the only rows in the matrix that do not use an API key.

The author treats the CLI result as the reason for calling the desktop behaviour a bug rather than something the Codex team is deliberately hiding, since the same API key reaches Fast mode from the command line without any patching at all.

The desktop client sends service_tier=None in place of service_tier=fast

With an API key selected, the desktop app exposes no Fast mode toggle at all, and edits to the Codex TOML configuration file have no effect on the application. In that setup there is no way to force Fast mode from the desktop client, according to the author's account of the behaviour.

The logs from the runs show the desktop app removing service_tier=fast from the request and passing service_tier=None. The fix was to modify the Codex desktop Electron app, published as codex-fast-mode-patcher on GitHub, with a warning that the patch is dangerous and should not be applied blindly by an agent.

What's next The author states he does not know why the desktop client removes the service tier, and the benchmark does not settle whether the behaviour is deliberate or accidental. The matrix covers only GPT-5.6-sol at high reasoning effort, leaving open whether the same stripping appears with other models or lower reasoning settings, and whether the patch keeps working across future desktop releases.