Codex Blocks Fast Mode with an API Key on Desktop

OpenAI's Codex Desktop app with an API key removes the Fast mode toggle and strips the `service_tier=fast` parameter, resulting in median throughput of 54.7 TPS versus 87.3 TPS when Fast mode is forced, according to benchmarks by developer adenta using the codex-tps tool with GPT-5.6-sol on high reasoning effort. The issue does not occur in the Codex CLI, which achieves 90.2 TPS with Fast mode, and adenta has published a patcher at github.com/adenta/codex-fast-mode-patcher to restore Fast mode in the Desktop app.

Ok, so you are Sonic. You gotta go fast. gotta go fast You recently switched to using an API key in the Codex Desktop app. You notice that there is no Fast mode toggle anymore. This bothers you. You start looking into this and things seem weird. When using an API key, it appears: - The fast mode toggle is no longer exposed in the UI. - Updating the Codex TOML config file changes nothing when using the Desktop application. You can’t force fast mode at all. - Fast mode with an API key works in the Codex CLI, but not in Desktop. The fact you can use fast mode with an API key in the CLI makes me think this is a bug vs something the Codex team is intentionally hiding. I built codex-tps https://github.com/adenta/codex-tps to benchmark this. I then ran a matrix of configuration options and measured their TPS responses. All experiments were run with GPT-5.6-sol on high reasoning effort. | Configuration | Median TPS | Branch and logs | |---|---|---| Stock Desktop + API key + service tier=fast | 54.7 | | 87.3 branch https://github.com/adenta/codex-tps/tree/experiments/desktop-api-green-forced-fast · log https://github.com/adenta/codex-tps/blob/b353c15f0ce687951d903095fb2753cbce26ad6e/experiments/codex-fast-mode-2026-08-13/logs/desktop-api-green-forced-fast--019ffb5c-eb09-75f0-8d1d-4d725de12858.jsonl branch https://github.com/adenta/codex-tps/tree/experiments/desktop-api-stock-tier-and-feature · log https://github.com/adenta/codex-tps/blob/b353c15f0ce687951d903095fb2753cbce26ad6e/experiments/codex-fast-mode-2026-08-13/logs/desktop-api-stock-tier-and-feature--019ffb67-abbd-7423-83fe-0139516b8e77.jsonl branch https://github.com/adenta/codex-tps/tree/experiments/desktop-api-stock-feature-only · log https://github.com/adenta/codex-tps/blob/b353c15f0ce687951d903095fb2753cbce26ad6e/experiments/codex-fast-mode-2026-08-13/logs/desktop-api-stock-feature-only--019ffb6c-796b-7fc1-b0c3-7c542c4d4139.jsonl branch https://github.com/adenta/codex-tps/tree/experiments/desktop-api-stock-baseline · log https://github.com/adenta/codex-tps/blob/b353c15f0ce687951d903095fb2753cbce26ad6e/experiments/codex-fast-mode-2026-08-13/logs/desktop-api-stock-baseline--019ffb71-0502-7873-8105-7e180b93f597.jsonl branch https://github.com/adenta/codex-tps/tree/experiments/cli-api-standard · log https://github.com/adenta/codex-tps/blob/b353c15f0ce687951d903095fb2753cbce26ad6e/experiments/codex-fast-mode-2026-08-13/logs/cli-api-standard--019ffb77-0db5-7f11-b02d-8f3bc474bded.jsonl 90.2 branch https://github.com/adenta/codex-tps/tree/experiments/cli-api-fast · log https://github.com/adenta/codex-tps/blob/b353c15f0ce687951d903095fb2753cbce26ad6e/experiments/codex-fast-mode-2026-08-13/logs/cli-api-fast--019ffb79-4fbb-75f0-a481-27dd90bdaf08.jsonl branch https://github.com/adenta/codex-tps/tree/experiments/desktop-chatgpt-standard · log https://github.com/adenta/codex-tps/blob/b353c15f0ce687951d903095fb2753cbce26ad6e/experiments/codex-fast-mode-2026-08-13/logs/desktop-chatgpt-standard--019ffb7f-4efb-7911-9e7d-3991e08a2099.jsonl 45.6 branch https://github.com/adenta/codex-tps/tree/experiments/desktop-chatgpt-fast · log https://github.com/adenta/codex-tps/blob/b353c15f0ce687951d903095fb2753cbce26ad6e/experiments/codex-fast-mode-2026-08-13/logs/desktop-chatgpt-fast--019ffb7f-8ab6-7ab3-90e5-d54687899dcb.jsonl As you can see, the stock Desktop app is indeed slower unless you manually go in and tweak the Codex Desktop code itself. If you dig into the logs, you can see that the Codex Desktop application is stripping out service tier=fast and passing service tier=None . I was able to fix this by updating the Codex Desktop electron app. I’m unsure why this is happening, but anyone else who wants to fix it can review this repo and patch their own Codex Desktop application like I did: github.com/adenta/codex-fast-mode-patcher https://github.com/adenta/codex-fast-mode-patcher . This is dangerous. Don’t blindly let your agent do this.