[FEATURE]: Support reasoning effort (thinking intensity) for xAI Grok 4.5

[MagicSpirit007]

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Summary

When using xAI Grok 4.5 in OpenCode (via xAI API / OAuth / @ai-sdk/xai), there is no way to adjust thinking / reasoning intensity (low / medium / high). The model effectively always runs at the provider default (high), and OpenCode does not expose effort variants for this model.

This matters for both latency and cost: Grok 4.5's default is high effort, so simple coding tasks burn more reasoning tokens than necessary if low/medium cannot be selected.

xAI API support

Per xAI Reasoning docs:

• grok-4.5 supports reasoning_effort / reasoning.effort
• Accepted values: low / medium / high
• Default: high
• Reasoning cannot be disabled (no none)

Also documented on the Grok 4.5 overview.

Example (chat completions style):

{
  "model": "grok-4.5",
  "reasoning_effort": "low",
  "messages": [{ "role": "user", "content": "…" }]
}

Current OpenCode behavior

In packages/opencode/src/provider/transform.ts, variants() only special-cases grok-3-mini for Grok reasoning effort:

// see: https://docs.x.ai/docs/guides/reasoning#control-how-hard-the-model-thinks
if (id.includes("grok") && id.includes("grok-3-mini")) {
  // only low / high
  return {
    low: { reasoningEffort: "low" },
    high: { reasoningEffort: "high" },
  }
}

grok-4.5 (and related Grok 4.x models that now expose effort) fall through and do not get low/medium/high variants in the model picker / Ctrl+T cycling / config.

reasoningEffort() already maps @ai-sdk/xai → { reasoningEffort: effort }, so the missing piece is mainly exposing the correct effort variants for Grok 4.5 (and wiring models.dev / capabilities if needed).

Expected behavior

For Grok 4.5 (and preferably other Grok models that support effort):

1. OpenCode exposes low / medium / high as selectable reasoning variants (same UX as other models with effort controls).
2. Selecting a variant sends the correct reasoningEffort / reasoning.effort to xAI.
3. Works for both xAI API key and Grok OAuth paths.
4. Do not offer none for Grok 4.5 (API rejects disabling reasoning).

Why this matters

Effort	Use case
low	Fast agent loops, simple edits, latency-sensitive work
medium	Balanced coding sessions
high (default)	Hard debugging / architecture / multi-step reasoning

Without this control, every request is effectively stuck on high, which is slower and more expensive than necessary for everyday agentic coding.

Related

• Prior request for an older model: [FEATURE]:Grok 4.3 reasoning levels #32459 ([FEATURE]: Grok 4.3 reasoning levels) — same class of feature; please extend coverage to Grok 4.5 with the official low|medium|high set (note: Grok 4.5 has no none, default is high).
• Older "enable thinking" request for Grok 4.2: [FEATURE]: Enable Grok 4.2 Thinking mode #17030 (closed) — different ask; this issue is specifically about adjustable effort levels, not merely turning thinking on.
• Token efficiency reports with Grok in OpenCode (e.g. Grok much more token efficient with grok build cli #38806) may also benefit if users can dial effort down when appropriate.

Suggested implementation sketch

1. Extend the Grok branch in ProviderTransform.variants() beyond grok-3-mini to cover grok-4.5 (and ideally a broader grok-4 / models.dev-driven set).
2. For Grok 4.5 specifically, expose:
   • low → { reasoningEffort: "low" }
   • medium → { reasoningEffort: "medium" }
   • high → { reasoningEffort: "high" }
3. For OpenRouter / compatible providers, use the matching { reasoning: { effort } } shape already used for other models.
4. Prefer driving this from models.dev reasoning_options if available, so future Grok models do not need one-off hardcoding.

Happy to test a PR once this lands.