admin The model, released on 12 August, scored 61 on the Artificial Analysis Intelligence Index, matching OpenAI’s GPT-5.6 Sol at its highest tested setting. Grok 4.6 improved by five points over Grok 4.5, which had scored 56, while Anthropic’s Claude Opus 5 remained ahead with 63 and Claude Fable 5 scored 62.
The result represents a substantial step for SpaceXAI as companies compete to develop models capable of handling prolonged, multi-stage assignments rather than simply answering individual prompts. Grok 4.6 has been designed particularly for long-running agents, software development, knowledge work and interactive or visual projects.
Its strongest results emerged in tasks requiring models to operate across several steps. Grok 4.6 posted an Elo score of 1,753 on the GDPVal-AA v2 knowledge-work evaluation, compared with 1,728 for GPT-5.6 Sol. On CursorBench 3.2, which measures coding performance, it scored 69.9%, against 67.2% for GPT-5.6 Sol and 70.5% for Claude Fable 5.
Performance was less uniform across every test. Grok 4.6 scored 65.9% on DeepSWE 1.1, below GPT-5.6 Sol’s 73% and Fable 5’s 70%. On FrontierCode 1.1 Extended, however, its 61.3% edged GPT-5.6 Sol’s 60.6%, while Fable 5 remained higher at 63.6%.
Independent evaluation also found Grok 4.6 particularly competitive on agentic workloads. It achieved 50.7% on a multi-turn banking benchmark involving customer-service tasks and tool use, while its Terminal-Bench v2.1 score of 88.4% placed it alongside leading models for software tasks performed through a terminal environment.
Price may become one of the model’s strongest competitive advantages. Grok 4.6 starts at $2 per million input tokens and $6 per million output tokens, unchanged from Grok 4.5. A faster version is available at twice those rates. The standard model remains substantially cheaper on headline API pricing than several models delivering comparable benchmark scores.
Measured costs reinforce that advantage. Grok 4.6 required about $0.84 per task in Artificial Analysis testing, placing it on the efficiency frontier among high-performing models. Its 500,000-token context window is smaller than the one-million-token window available with GPT-5.6 Sol, creating a trade-off for workloads involving exceptionally large documents or long running contexts.
SpaceXAI extended the supplemental training used for Grok 4.6 compared with its predecessor. The process included curated model-generated reasoning data, engineering material and modifications to the training optimisation process. Supervised fine-tuning was followed by reinforcement learning across knowledge work, general coding, kernel optimisation, web development and computer-aided design.
The company has also placed greater emphasis on models capable of producing complete applications and work products. During internal testing, Grok 4.6 was tasked with researching unfamiliar areas, structuring applications, implementing core functions and refining results through several rounds of feedback. SpaceXAI said longer runs showed more instances of the system testing and checking its own work before proceeding.
That direction reflects a broader shift across frontier AI development. OpenAI has been strengthening GPT-5.6 Sol for complex research, coding and professional workflows while also improving its factual reliability and consistency in ChatGPT. An August update brought the same Sol model across faster responses and deeper reasoning modes for Plus and Pro users.
