TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
xAI Challenges GPT-5.6 with Grok 4.6, Long-Horizon Agents Will Self-Check Work

xAI Challenges GPT-5.6 with Grok 4.6, Long-Horizon Agents Will Self-Check Work

2026.08.13
Share
Grok 4.6 focuses primarily on long-task agents and visual project capabilities, matching GPT-5.6 Sol on the Comprehensive Intelligence Index while maintaining highly competitive API pricing.

Author: xAI

Compiled by: TechFlow

TechFlow Insight: Grok 4.6 focuses on agents capable of long-running, multi-step work, and ties with GPT-5.6 Sol on the Composite Intelligence Index, offering practical reference value for developer selection. xAI also offers double usage quota in the first week, indicating that large model competition is shifting from single-turn Q&A to agent capabilities capable of actually delivering projects, worth tracking for practitioners concerned with AI and crypto integration opportunities.

Today we are releasing Grok 4.6. Grok 4.6 is built on Grok 4.5, with a special focus on long-running agents and more ambitious interaction and visual work. It can continuously handle complex tasks across multiple steps, whether researching topics, analyzing information, working across codebases, or turning ideas into polished applications or creations.

Grok 4.6 reaches frontier levels on multiple agent coding and knowledge work benchmarks. It ties with GPT-5.6 Sol on the Artificial Analysis Intelligence Index, which is a composite score of nine benchmarks.

Figure: Source: xAI; Competitor data taken from system cards or benchmark leaderboards publicly released by various developers.

Grok 4.6 is available today in Cursor and Grok Build. During the first week, we are offering 2x included usage in Grok Build and Cursor to make it easy for you to start trying out 4.6 immediately.

Training Grok 4.6

Grok 4.6 underwent longer supplemental training than Grok 4.5, using filtered model-generated data for reasoning and advanced technical concepts, as well as high-quality engineering data, and improved optimizers and training recipes. This laid a stronger foundation for subsequent SFT and RL stages.

Subsequently, we used Grok 4.5 to regenerate SFT trajectories, covering reasoning work, agent frameworks, and areas such as STEM, software engineering, and knowledge work, and filtered out problematic trajectories via model-based checks. The resulting SFT checkpoints exhibited strong performance and better behavior.

Grok 4.6 was trained on extensive agent reinforcement learning tasks, including knowledge work, general coding, and specific environments for areas such as kernel optimization, web development, and computer-aided design.

Turning Grand Ideas into Usable Projects

We tested Grok 4.6 on projects aimed at expanding its scope and long-duration work capabilities. We found that the model is particularly adept at turning broad product ideas into usable first versions. It can research unfamiliar domains, build application structures, implement core interactions, and continuously optimize results through multiple rounds of feedback.

In longer task trajectories, we also began to see more self-testing and validation behaviors, where the model checks its own work before proceeding.

On visual and interaction projects, Grok 4.6's initial output is stronger than what we typically see on Grok 4.5. Given a concrete product idea, it can establish structure and visual language for an application in one go. This makes it particularly useful in projects where the fastest way to get good results is to create substantial content first and then iterate.

Safety and Capabilities

Grok 4.6's safety guardrails have been improved and calibrated according to model capabilities.

Our safety stack aims to maximize utility and safety in legitimate use cases, making Grok 4.6 both useful and safe in areas such as vulnerability patching, accelerating engineering design cycles, and enhancing AI research.

Our safety guardrail evaluation work reflects Grok 4.6's expanded capabilities, including the most extensive capability and safety guardrail calibration pre-deployment testing we have ever conducted, as well as extensive post-deployment and third-party testing.

Benchmark Evaluation

Figure: Comparison of scores between Grok 4.6 and competitors on major benchmarks. Source: xAI; The highest score for each evaluation is shown in bold, third-party model scores are taken from their self-reported or publicly available best results.

From public data, Grok 4.6 ties or exceeds GPT-5.6 Sol Max on CursorBench v3.2 (69.9%), GDPVal-AA v2 (1753), AA-Briefcase (1577) and APEX-Agents (57.5%), but still lags on Terminal-Bench v3.0 (26%) and DeepSWE v1.1 (65.9%); Fable 5 Max maintains a slight lead on most items.

Getting Started with Grok 4.6

Grok 4.6 is available today in Cursor and Grok Build. It is also available via API and other partners, such as OpenRouter, Vercel, and Cloudflare.

Pricing starts at $2 per million input tokens, $6 per million output tokens. Additionally, there is a fast version, priced at double the standard version.

During the first week, we are offering 2x included usage in Grok Build and Cursor to make it easy for you to start trying out 4.6 immediately.

Create API Key

Start building with Grok 4.6 via the xAI API immediately.

Read the documentation and integrate Grok 4.6 into your tech stack.

Try for Free in Grok Build

Go to x.ai/build immediately to start using.