Codex OAuth Fast models run at standard throughput

[AidenGeunGeun]

Description

On vanilla OpenCode, ChatGPT-authenticated Fast model entries send service_tier: "priority", but do not show the speed increase that the same account gets in the official Codex CLI.

Tested with clean upstream checkouts, no plugins, and --pure:

• v1.18.10: HTTP and Responses WebSocket
• current dev (14f0bf64a19493110b51f5fdeb9c1c1bba5dd3f5): Responses WebSocket
• gpt-5.6-sol vs gpt-5.6-sol-fast, low effort
• identical fixed prompt producing 603 output tokens

Current dev, 3 alternating runs per model (time from OpenCode step_start to step_finish):

Model	Median turn time	Median output rate
gpt-5.6-sol	10.818s	55.74 tok/s
gpt-5.6-sol-fast	10.950s	55.07 tok/s

The earlier clean v1.18.10 test was also flat on both transports. In contrast, official codex-cli 0.145.0 on the same account and prompt completed Standard in a median 16.30s and Fast in 11.11s across 4 alternating runs each (1.47x).

A temporary redacted probe at OpenCode's OAuth fetch boundary confirmed the current upstream request behavior:

standard: model=gpt-5.6-sol, service_tier omitted
fast:     model=gpt-5.6-sol, service_tier=priority

I am not treating response.service_tier=default as proof because OpenAI has stated that field is not a reliable end-to-end Fast indicator for ChatGPT auth: openai/codex#14204 (comment). The controlled speed result and same-account Codex CLI control are the relevant signals here.

Related but not duplicate: #23267 and #16499 reported Fast feeling no different, but were closed once a Fast model appeared in the picker; runtime routing was not verified. #25059 likewise confirms the catalog design, not the serviced speed.

Could maintainers clarify the sanctioned way for third-party ChatGPT OAuth clients to request Codex Fast routing? If OpenCode needs Codex-specific metadata or another integration beyond service_tier, where should that live? I would be happy to work on a PR once the desired direction is confirmed. If proper routing is not currently available to third-party clients, not advertising Fast for this auth path may be a safer interim behavior.

Plugins

None (--pure)

OpenCode version

v1.18.10 and current dev at 14f0bf64a19493110b51f5fdeb9c1c1bba5dd3f5

Steps to reproduce

1. Sign in to OpenAI using ChatGPT OAuth.
2. From a clean upstream build, alternate these commands several times with the same prompt:

opencode run --pure --format json --model openai/gpt-5.6-sol --variant low   "Output exactly the integers from 1 to 300 inclusive, one per line, with no commentary."
opencode run --pure --format json --model openai/gpt-5.6-sol-fast --variant low   "Output exactly the integers from 1 to 300 inclusive, one per line, with no commentary."

3. Compare step_start to step_finish timestamps and verify each response has 603 output tokens.
4. As a control, run the same prompt in Codex CLI with Fast off/on.

Operating System

macOS 27.0, Apple Silicon

Terminal

zsh

[davhdavh]

I can confirm that speed is the same in my testing also for openai/gpt-5.6-luna and openai/gpt-5.6-luna-fast