[App] Advanced power slider skips Max and mixes model, reasoning, and Ultra #31968

What version of the Codex App are you using (From “About Codex” dialog)?

Codex App 26.707.31428 (codex-cli 0.144.0-alpha.4)

What subscription do you have?

Not plan-specific. Reproduced on an account with access to GPT-5.6 Sol, Max, and Ultra.

What platform is your computer?

Darwin 24.6.0 arm64 arm

What issue are you seeing?

The Advanced “power” slider looks like one continuous intelligence/reasoning scale, but its stops mix two different dimensions (model choice and reasoning effort) and silently omit max.

The effective preset sequence in this build is:

5.6 Terra + low
5.6 Sol   + low
5.6 Sol   + medium
5.6 Sol   + high
5.6 Sol   + xhigh
5.6 Sol   + ultra

There is no 5.6 Sol + max stop. Moving the slider right from “5.6 Sol Extra High” therefore jumps directly to “5.6 Sol Ultra”, even though:

• GPT-5.6 supports max as a distinct reasoning effort.
• max can appear in the full reasoning-effort dropdown when enabled.
• Ultra has separate UI treatment and a “Consumes usage limits faster” warning.
• OpenAI’s model guide describes max as the hardest quality-first reasoning effort, while Ultra is a Codex product mode with different orchestration semantics.

This is extremely hard to reason about as a user. Is the slider choosing a model, a reasoning budget, an orchestration mode, or all three? At the moment the honest answer is “all three, except one of the supported levels is skipped.”

Why this setting is so confusing

This is more than a missing slider stop. The visual design strongly implies one simple, monotonic “more intelligence” scale, but moving a single control can silently:

• switch from Terra to Sol,
• change the reasoning effort,
• skip a supported effort (max), and
• finally switch into the special Ultra execution mode.

Meanwhile, the full Advanced dropdown presents a different ladder, and the Simplified Chinese UI can render two distinct modes with the same name. A normal user therefore cannot tell what capability they selected, why Max disappeared, whether Ultra is “more reasoning” or a different execution strategy, or how the choice affects usage limits.

This setting should be understandable from the UI itself. Users should not have to inspect the bundled application code or read API documentation just to discover what each slider position actually changes.

The Simplified Chinese locale makes the confusion worse because xhigh and ultra are both translated as 极高; that separate deterministic localization bug is already tracked in #31963. This issue is specifically about the slider/preset model and the missing max stop.

What steps can reproduce the bug?

1. Open the Codex App with access to GPT-5.6 Sol and Ultra.
2. Expand Advanced above the composer.
3. Move the power slider to 5.6 Sol Extra High.
4. Move it one stop to the right.
5. Observe that it becomes 5.6 Sol Ultra; there is no 5.6 Sol Max stop.
6. Open the full model/reasoning controls and compare them with the slider. When max is enabled under Settings → General → Model features → Available reasoning efforts, it is a valid selectable effort there but remains absent from the shortcut slider.

What is the expected behavior?

The control should expose a coherent mental model. Any of these would be substantially clearer:

1. If the slider is an ordinal reasoning-effort control, keep the model fixed and include max between xhigh and ultra.
2. If it is intentionally a curated cross-model “power” preset control, disclose that each stop may change both model and execution mode, and explain why max is omitted.
3. Treat Ultra as a separate toggle/mode rather than placing it as an unexplained ordinal successor to Extra High.

At minimum, the full dropdown and the shortcut slider should not imply contradictory effort ladders.

Additional information

• Current GPT-5.6 guidance lists none, low, medium, high, xhigh, and max as supported reasoning efforts and says to reserve max for the hardest quality-first workloads: https://developers.openai.com/api/docs/guides/latest-model
• The installed app recognizes both max and ultra as distinct values and exposes a “Available reasoning efforts” setting, but the current slider preset sequence explicitly jumps from xhigh to ultra.
• Related localization bug: [App][i18n] zh-CN renders both xhigh and ultra reasoning efforts as “极高” #31963.