Grok 4.7

What is disclosed

Grok 4.7 reached the xAI API on 21 September 2026.¹¹1xAI, API release notes, entries of 21 September and 12 August 2026. https://docs.x.ai/developers/release-notes. Arithmetic: $4, $1 and $12 above 200K against $2, $0.50 and $6 below, each 2 times. xAI, which now trades as SpaceXAI, calls it its frontier model for coding and knowledge work.²²2xAI, Grok 4.7 model card, 21 September 2026, pp. 3-4, which also states that SpaceXAI is a doing-business-as name of XAI LLC. https://media.x.ai/v1/website/4p7card-5eccc980.pdf ³³3xAI, Grok 4.7 overview, docs.x.ai, last updated 28 September 2026. https://docs.x.ai/developers/grok-4-7 The announcement gives one fact about its size: Grok 4.7 "uses a new, larger base model" than Grok 4.6.⁴⁴4xAI, Introducing Grok 4.7, 21 September 2026, read in the Internet Archive copy of 29 September 2026. https://x.ai/news/grok-4-7 and https://web.archive.org/web/20260929101703/https://x.ai/news/grok-4-7 The parameter count and the architecture are not published.

The context window is 500,000 tokens.³³ xAI sets no limit on text output. Images can go in, but only text comes out.³³

The docs give May 2026 as the knowledge cutoff.³³ The model card says supplemental training used data from as late as August 2026, some of it anonymized workflow data from Cursor.²²

Reasoning effort has four levels from low to xhigh. The default is high.³³

What is new

Grok 4.7 keeps the context window and the prices of Grok 4.6, which came out on 12 August 2026.¹¹ The reinforcement learning run was longer and used harder tasks, weighted toward problems that take many hours to complete.⁴⁴

xAI reports 46.3% on CursorBench 4.0 at xhigh effort, where Grok 4.6 scored 40.4% at high.⁴⁴ On DeepSWE v1.1, at high effort, it reaches 71.0%.⁴⁴

The safeguards were rebuilt from scratch. On HackerBench v0.3, xAI's own test of risky and malicious cyber tasks, 3.3% of risky dual-use prompts got through.⁴⁴ Selected security partners have invite-only access to the model's red-team capabilities.⁴⁴

Using it

Requests go to grok-4.7, on the Responses API or on Chat Completions.³³ Below 200K prompt tokens, input costs $2 per million tokens and output $6, with cached input at $0.50.¹¹ Past that line all three rates double.¹¹ The higher rates then apply to every token in the request.⁵⁵5xAI, API pricing, docs.x.ai, read 29 September 2026. https://docs.x.ai/developers/pricing The Batch API does not take the model.⁶⁶6xAI, Grok 4.7 model page, docs.x.ai, read 29 September 2026. https://docs.x.ai/developers/models/grok-4.7

On the Responses API every reply carries the model's reasoning in encrypted form. Clients pass it back unchanged in the next request.³³ The docs also warn that requests sent without a prompt cache key often land on a server with a cold cache and pay the full input price.³³

A US regional endpoint keeps inference in the United States, at a 10% premium on tokens.³³ Grok 4.7 Fast is the same model on faster infrastructure, at twice the standard rates. It is sold only inside Cursor and Grok Build, not on the public API.³³

Grok 4.3 holds twice the context at lower prices.⁷⁷7xAI, Grok 4.3 model page, docs.x.ai, read 29 September 2026: 1,000,000 tokens of context, $1.25 input and $2.50 output. https://docs.x.ai/developers/models/grok-4.3. Arithmetic: 1,000,000 / 500,000 = 2.