xAI's current flagship, superseding Grok 4.5 at the same headline rate: $2 in / $6 out per million below a 200K prompt, doubling above it. A 500K context window and a February 2026 knowledge cutoff.
Why this oneWhat it is actually for
Reach for grok-4.6 when the job needs current or social context — xAI's differentiator is real-time knowledge through X, and the model page recommends it for code, chat and general work as "the most intelligent and fastest model we've built".
Choosing it over Grok 4.5 costs nothing: the rate card is identical at both prompt tiers. 4.5 remains listed and this board keeps it, because a comparison that silently drops a model a reader may still be running is worse than one row too many.
What it isArchitecture, lineage, training
Grok 4.6 is a closed-weight model with a 500K-token context window. Its pricing is tiered by prompt length, which is the single most important thing to know before costing it: everything below 200K input tokens bills at one rate and everything at or above bills at double. A 500K window on a model that doubles at 200K is a window you should enter deliberately.
Its knowledge is fixed at 1 February 2026. Without search enabled it has no awareness of anything after that date — which matters more here than on most models, because real-time knowledge is the reason to pick it.
At a glanceSee it
The 200K line is the whole cost story on Grok 4.6 — crossing it doubles both input and output, and the 500K window makes that easy to do by accident.
CapacityContext, output and what fits
| Fact | Value | Source |
|---|---|---|
| Model | Grok 4.6 | xAI models page |
| Context window | 500K tokens | xAI models page |
| Knowledge cutoff | 1 February 2026 | xAI models page |
| Pricing tier boundary | 200K input tokens | xAI models page |
| Cached input | $0.50 per M below 200K; $1.00 at or above | xAI models page |
| Max output | Not documented on the vendor page | — |
| Vision and tool support | Not documented on the models page read here | — |
The parametersEvery knob, and what moving it does
| Parameter | What it does | Range or default | What happens when you move it |
|---|---|---|---|
| Prompt length | Selects the billing tier | Under or over 200K input tokens | Crossing 200K doubles BOTH input and output for the whole request |
| Cached input | Reuses a prefix | $0.50 per M under the line | Four times cheaper than a miss at the same tier |
| Live search | Brings in post-cutoff knowledge | Off unless enabled | Without it the model knows nothing after 1 Feb 2026 |
SamplingShaping the output distribution
The xAI models page read for this entry does not document sampling controls for Grok 4.6. This page does not restate the Grok 4.5 behaviour as though it still held; check the vendor's API reference before relying on a specific range.
ReasoningThinking, effort and budgets
The models page does not document reasoning-effort controls for Grok 4.6, and no figure is asserted here. What it does state is the positioning — "the most intelligent and fastest model we've built" — and a recommendation for code, chat and general use.
ToolsFunction calling and server tools
Search is the capability xAI documents and the one that defines the model: without it, knowledge stops at 1 February 2026. Function calling and structured-output support are not described on the models page read here, so no claim is made.
CostPrice, caching, batching, what drives the bill
List price from xAI's models page, per million tokens, at the tier this board quotes (below 200K input tokens):
| Line | Under 200K | 200K and above |
|---|---|---|
| Input | $2.00 | $4.00 |
| Cached input | $0.50 | $1.00 |
| Output | $6.00 | $12.00 |
The board states the short tier because that is where the reference workload sits. Anyone pricing long-context work must double both figures.
Where it runsSurfaces and availability
| Surface | Available | Notes |
|---|---|---|
| xAI API | Yes | The documented route |
| Open weights | No | Closed API |
| Fine-tuning | Not documented | The models page read here does not describe it |
StrengthsWhat it is good at
- xAI's current flagship, and recommended by the vendor over 4.5 for code, chat and general work.
- Real-time knowledge through search — the reason to choose this family at all.
- Identical headline price to the model it supersedes, so upgrading costs nothing.
- A 500K window at the same rate as smaller prompts, up to the 200K line.
LimitsWhere it falls down
- Price doubles at 200K input tokens, on both input and output.
- Knowledge stops at 1 February 2026 without search.
- Max output, vision, function calling and structured outputs are not documented on the models page, so none is claimed here.
- A smaller tooling ecosystem than the OpenAI and Anthropic families.
Against its neighboursHow it compares
Against grok-4.5: same price at both tiers, same 500K window, and the vendor recommends 4.6. 4.5 is kept on this board because 202 measured kits cite it and dropping a row would silently change what those comparisons mean.
Against claude-sonnet-5 at $2/$10: the same input rate and a lower output rate here ($6 against $10), with a smaller window (500K against 1M) and a documented ecosystem gap.
Getting startedThe smallest call that works
Measure your prompt length distribution first. On this model that is not a performance question, it is the billing tier — and a pipeline whose prompts sit near 200K will see its bill double without anything in the code changing. Cache the stable prefix; the hit rate is four times cheaper than the miss at either tier.
SourcesWhere every claim above came from
- xAI models and pricing —
docs.x.ai/docs/models(read 18 Sep 2026)
Price and capacity verified 2026-09-18 against https://docs.x.ai/docs/models. first entry 2026-09-18 (primary source read today: https://docs.x.ai/docs/models — grok-4.6 at $2.00 per million input and $6.00 per million output below 200k prompt tokens, $4.00/$12.00 at or above. The same read confirmed grok-4.5 remains listed at an unchanged $2.00/$6.00, so the existing row's price is correct and was re-stamped rather than altered.)