# xAI reasoning — always-on reasoning models, `reasoning_effort`, `reasoning_content`, encrypted reasoning

**Status:** `DOCUMENTED` + `LIVE_VERIFIED` (2026-09-19). Machine-readable: `parameters/xai-chat-completions.json` (`reasoning_effort`), `parameters/xai-responses.json` (`reasoning.*`, `include`), `objects/xai-inference-objects.json` (Reasoning item).

**Sources:** https://docs.x.ai/developers/model-capabilities/text/reasoning · https://docs.x.ai/developers/model-capabilities/text/generate-text#returning-encrypted-thinking-content · https://docs.x.ai/developers/model-capabilities/text/multi-agent · https://docs.x.ai/developers/models (live `/v1/language-models`)
**Last verified:** 2026-09-19

## Model classes (live `GET /v1/language-models`, 7 text models)

| Model | Reasoning | `reasoning_effort` values | Live notes |
|---|---|---|---|
| `grok-4.6` (aliases none) | always on | `low` `medium` `high` (default) `xhigh` | knowledge cutoff 2026-02-01; summarized reasoning exposed |
| `grok-4.5` (`grok-4.5-latest`, `grok-build-latest`) | always on | `low` `medium` `high`; `xhigh` treated as `high` | |
| `grok-4.3` (`grok-4.3-latest`) | always on | live: `low`/`medium`/`high`/`xhigh` all 200; **`none` → 200 with `reasoning_tokens: 0` (LIVE_DISCOVERED, undocumented)** | 70–180 reasoning tokens on "Reply with OK."; response echo `reasoning.effort:"low"` when unset |
| `grok-4.20-0309-reasoning` (`grok-4.20`, …) | always on | docs: see model page | |
| `grok-4.20-0309-non-reasoning` (`grok-4.20-non-reasoning`, …) | **none** | ❌ `reasoning_effort` → 400 "does not support parameter reasoningEffort" | `reasoning_tokens: 0`, no `reasoning_content` key; Responses echo `reasoning:{effort:null, summary:null}` |
| `grok-4.20-multi-agent-0309` | multi-agent | `effort` controls agent count (low/medium → 4, high/xhigh → 16) | not called |
| `grok-build-0.1` (`grok-code-fast-1`, …) | reasoning (coding) | | not called |

Reasoning **cannot be disabled** per docs (except via the live `none` value on grok-4.3). `presence_penalty`, `frequency_penalty`, `stop` return 400 on reasoning models.

## Where the reasoning shows up

| Surface | Field | Content |
|---|---|---|
| Chat Completions | `choices[].message.reasoning_content` (string) | short trace/summary, e.g. `The user requested a reply of "OK."`; streamed as `delta.reasoning_content` before content |
| Chat usage | `usage.completion_tokens_details.reasoning_tokens` | billed at output price; `total_tokens` includes them |
| Responses | `output[]` item `{type:"reasoning", id:"rs_…", summary:[{type:"summary_text", text}], status, encrypted_content?}` | summary always "detailed" per docs; item sometimes omitted; `usage.output_tokens_details.reasoning_tokens` |
| Responses stream | `response.reasoning_summary_part.added/…text.delta/…text.done/…part.done` | `response.reasoning_text.delta` documented, not observed |
| `/v1/messages` | `content[] {type:"thinking", thinking, signature:""}` (+ stream `thinking_delta`) | appeared with tool_use and in streams; `usage.output_tokens` includes reasoning |

## Encrypted reasoning (stateless multi-turn / ZDR)
`include: ["reasoning.encrypted_content"]` (or xai-sdk `use_encrypted_content=True`) adds `encrypted_content` to the reasoning item (also visible in the stream's `output_item.done`). Replay `...response.output` in the next `input` — live 200. Required for cache hits and correct agentic state when not using `previous_response_id`. Also the mechanism behind context compaction. Vercel AI SDK does this automatically unless `store:false`.

## Cost
Reasoning tokens bill at the output rate (grok-4.3: $2.50/M). A minimal call ≈ 200 prompt + 2 output + ~130 reasoning ≈ $0.0006 (`usage.cost_in_usd_ticks` 3 809 000 = $0.00038 with cache). Batch API discounts (20 % on grok-4.3 / 4.20) apply to reasoning tokens too.