[Digital Today intern reporter Seungah Yoo (유승아)] SpaceXAI unveiled its new artificial intelligence (AI) model, Grok 4.6.

Foreign media including IT Media reported on Aug. 12 local time that the model posted the same score as OpenAI's GPT-5.6 Sol Max on a composite performance index. It focuses on agent use that continues tasks across multiple steps and on generating conversational and visual outputs. SpaceXAI is an AI organisation under SpaceX.

Grok 4.6 was developed based on the previous generation, Grok 4.5. SpaceXAI said it raised performance for agent use that continuously carries out multi-step tasks and also put emphasis on creating visual outputs within conversations with users.

In benchmarks presented by the company, Grok 4.6 scored 61 in Artificial Analysis' Intelligence Index (AAII), which sums nine metrics. That is the same level as GPT-5.6 Sol Max. It was 1 point short of the 62 scored by Anthropic's Fable5 Max. It was 5 points higher than the 56 scored by the previous model, Grok 4.5 High.

Strengths and weaknesses were mixed by category. In GDPVal-AA v2, which evaluates the ability to perform work with economic value, GPT-5.6 Sol Max scored 1,753. In AA-Briefcase, which assesses knowledge work, it scored 1,577. In Harvey Lab (Vals), a legal evaluation, it scored 15.8 percent, the highest among the comparison models.

By contrast, Grok 4.6 led Fable5 Max and Grok 4.5 High after scoring 65.9 percent in DeepSWE v1.1, a software engineering evaluation, and 26 percent in Terminal Bench v3.0, which measures terminal operation performance. SpaceXAI said it cited competitors' figures from system cards and leaderboards released by each developer.

The company said Grok 4.6 is particularly strong at creating a first version that works from only a rough product idea. It said the model is suited to researching unfamiliar fields, designing an application architecture and iterating improvements by incorporating feedback. It added that during long tasks, the model has begun to show behaviour in which it verifies outputs on its own before moving to the next step.

It also revised training methods. SpaceXAI applied longer pre-training than Grok 4.5 and conducted supervised fine-tuning (SFT) using data regenerated by Grok 4.5. It added reinforcement learning that includes domain-specific environments such as kernel optimisation, web development and computer-aided design (CAD).

Safety reinforcement was also carried out. The company said it improved and adjusted safeguards in line with performance gains and that the scale of pre-release performance evaluations and safeguard adjustment tests was the largest so far.

It also expanded distribution channels. Grok 4.6 is available from the day of release via Grok Build, Cursor, Grok Bot and an API. It is also available through partner channels such as OpenRouter, Vercel and Cloudflare. Among them, the Grok Bot beta is a service SpaceXAI unveiled the previous day and is an AI agent that runs continuously on a virtual computer in the cloud. It can be used on iPhone and Mac. SpaceXAI previously acquired Cursor in June.

Pricing starts at $2 per 1 million input tokens and $6 per 1 million output tokens. A high-speed version costs twice as much. SpaceXAI is also running a promotion for 1 week after launch that doubles the plan-included usage on Cursor and Grok Build.