SpaceXAI ships Grok 4.6: a post‑training update, unchanged pricing and a new xhigh reasoning level

On 12 August 2026 SpaceXAI (formerly xAI) published the announcement of Grok 4.6 on its official blog, the direct successor to Grok 4.5. The model is available immediately on Cursor and Grok Build, through the API on console.x.ai, and via partners OpenRouter, Vercel and Cloudflare, with double the included usage for the first week (source: SpaceXAI, official announcement). The API documentation lists USD 2 per million input tokens, USD 0.50 per million cached input tokens and USD 6 per million output tokens up to 200,000 prompt tokens; beyond that threshold the rates double (USD 4 / 1 / 12). The “fast” variant costs twice the standard rate. The context window is 500,000 tokens, unchanged from Grok 4.5, and the model accepts text and images as input but generates text only (source: SpaceXAI, API release notes).

SpaceXAI lists four reasoning levels – low, medium, high (the default) and the new xhigh – and publishes a set of benchmarks for the High configuration: Artificial Analysis Intelligence Index 61, GDPval‑AA v2 1753 Elo, CursorBench v3.2 69.9%, DeepSWE v1.1 65.9%, FrontierCode v1.1 61.3%, APEX‑Agents 57.5%, APEX‑SWE 56.4%, AA‑Briefcase 1577, Harvey LAB 15.8% and Terminal‑Bench v3.0 26% (source: SpaceXAI, official announcement). The independent body Artificial Analysis confirms the intelligence index of 61, the USD 2 and USD 6 per million token prices and the width of the window, adding a generation speed of 65.5 tokens per second, below the median for comparable models (source: Artificial Analysis, independent profile).

VentureBeat reports that with these results Grok 4.6 overtakes Kimi K3 and draws level with GPT‑5.6 Sol, taking third place in the Artificial Analysis ranking. Every benchmark except the Artificial Analysis index, however, comes from the vendor and has not been checked by independent third parties. The parameter count (around 1.5 trillion) is quoted by some aggregators, but appears in no official source and cannot be verified (source: unspecified aggregators). The knowledge cutoff date (1 February 2026) likewise comes from unofficial sources (MarkTechPost) and is not stated in SpaceXAI’s documentation. And nobody has independently measured what the doubled rates beyond 200,000 prompt tokens actually mean for long agentic workloads, leaving the real cost of heavy use an open question.

On Terminal‑Bench v3.0, which measures command‑line software engineering, Grok 4.6 stops at 26% against the 34% credited to GPT‑5.6 Sol Max: by SpaceXAI’s own tables, that is the model’s weakest point.

SpaceXAI credits the performance jump to extra training on curated data generated by the model itself, an improved optimiser and agentic reinforcement learning across multiple domains – not to a larger base model. No weight release is planned: the model stays accessible only through the API and partner platforms.

Grok 4.6 arrives roughly five weeks after Grok 4.5 with the price list untouched. In a year when the comparison between frontier models has shifted from absolute score to the intelligence‑per‑dollar ratio – especially on agentic workloads, where a single task can burn through millions of tokens – the positioning is that of a cheap alternative to models with the same score.

In short, Grok 4.6 is an incremental post‑training update, with clear gains on some benchmarks (DeepSWE, APEX‑Agents, GDPval‑AA v2, AA‑Briefcase) and a pricing policy unchanged from Grok 4.5. The lack of independent verification for most of the results, the absence of long‑term cost data and the Terminal‑Bench weak spot all argue for caution in the overall assessment.