Grok 4.6

Grok 4.6 is xAI’s current generation, released 12 August 2026. Its announcement and model card both attribute it to SpaceXAI rather than to xAI — the company was acquired by SpaceX in February 2026 and renamed in July, and this is the model where the new name appears on the vendor’s own pages.

No architecture is published. No parameter count, no layer count, no expert configuration. Grok-1 in 2024 was the last generation xAI described, and everything since has been a product announcement rather than a technical one.

What xAI does describe: where the compute went#

The disclosure that exists is about training emphasis rather than structure, and it is worth reading as such.

Grok 4.6 is built on Grok 4.5. Its supervised fine-tuning trajectories were regenerated by Grok 4.5 — the previous model producing the worked examples the next one learns from — supplemented with curated model-generated reasoning data and engineering datasets. On top of that sits reinforcement learning across agentic tasks spanning knowledge work, coding and domain-specific environments.

Two things follow. First, the generation-to-generation improvement is being driven substantially by the previous model’s own output, not by new human data. Second, “trained with reinforcement learning at scale” is a claim about budget allocation, not about what the model is made of — the same kind of claim OpenAI makes about its reasoning models, and equally uninformative about architecture.

Reasoning is exposed as four effort levels: low, medium, high (the default) and xhigh.

X Search, the actual differentiator#

Grok’s distinguishing capability is not architectural. It is access to posts on X, the platform its corporate parent owns.

The documentation describes X Search as a capability with four modes — keyword search, semantic search, search by user, and fetching a specific thread — with handle and date filters, and understanding of images and video inside posts. It is billed per call, separately from tokens.

No other builder in this section has this, and none can acquire it by training a better model; it follows from common ownership of a large social platform. Whether it is an advantage depends entirely on whether posts on X are a good source for the question at hand — which is a question about the corpus, not about the model.

A documentation trap worth knowing: docs.x.ai serves two trees, and the older one describes only ordinary web search. It answers confidently rather than erroring, so it is easy to conclude from it that the X integration does not exist.

Positioning, and what comes with it#

xAI’s public stance is that its models refuse less than competitors’. That is a product decision rather than a technical property, and the builder page carries the four documented episodes that have followed from it — two of which xAI attributed to changes in the system prompt, one of those unauthorised.

The durable consequence is one worth stating plainly: after the second episode, xAI began publishing its system prompts on a public repository. Whatever else is true of the sequence, that is a transparency practice no other builder in this section matches, and it came out of a failure rather than a policy.

Status#

Checked 11 September 2026. xAI’s model documentation is the live source — check that a page is under the current documentation tree before relying on it.

Grok 4.6 is the current generation, released 12 August 2026. Grok 4.5 (16 July 2026) and Grok 4.3 sit behind it and remain available. A multi-agent “Heavy” tier runs several instances against a problem and combines the results.

Access is through xAI’s own interface, through X, and — since SpaceXAI’s acquisition of Anysphere in August 2026 — as a first-party surface in the Cursor editor. Image and video generation ship separately under the Grok Imagine brand.

Sources#

Backward xAI (Grok) Grok-1 Forward