xAI has launched Grok 4.6, a new version built on Grok 4.5 with a clear emphasis on long-running agents, richer interactive workflows and visual project creation. The company said the model is designed to keep working through complex tasks over multiple steps, including topic research, information analysis, work across codebases, and turning ideas into completed applications or creative output.
xAI said Grok 4.6 reaches frontier-level results on several agentic coding and knowledge-work benchmarks and matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a composite score built from nine benchmarks.
Available now in Cursor and Grok Build
Grok 4.6 is available today in Cursor and Grok Build. For the first week, xAI said it is offering 2x included usage in both products so users can start testing the model right away.
Longer post-training and updated data pipeline
According to xAI, Grok 4.6 went through longer supplementary training than Grok 4.5. The company used curated model-generated data for reasoning and advanced technical concepts, added high-quality engineering data, and changed the optimizer and training recipe. xAI said those steps created a stronger base for later supervised fine-tuning, or SFT, and reinforcement learning, or RL.
xAI then regenerated SFT trajectories with Grok 4.5 across reasoning tasks, agent frameworks, STEM, software engineering and knowledge work, while filtering problematic trajectories with model-based checks. The resulting SFT checkpoint showed stronger performance and improved behavior, according to the company.
Grok 4.6 was also trained on a broad set of agentic reinforcement learning tasks spanning knowledge work, general coding, and domain-specific environments such as kernel optimization, web development and computer-aided design.
Project building and self-verification on longer tasks
xAI said it tested Grok 4.6 on projects intended to stretch its scope and long-duration working ability. In those tests, the model was said to be particularly strong at turning broad product ideas into usable first versions. The company said it can research unfamiliar domains, build application structure, implement core interactions and keep improving results through repeated feedback cycles.
On longer task trajectories, xAI said it is starting to see more self-testing and verification behavior, with the model checking its own work before continuing.
For visual and interactive projects, xAI said Grok 4.6 produces stronger first-pass output than it typically saw with Grok 4.5. Given a specific product idea, the model can set up an app’s structure and visual language in one shot. xAI said that makes it useful for projects where the fastest path to a good result is to build substantial content first and then iterate.
Safety stack updated alongside capability gains
xAI said Grok 4.6’s safeguards were improved and calibrated in line with the model’s capabilities. The company said its safety stack is meant to maximize both utility and safety for legitimate use cases, including vulnerability patching, faster engineering design cycles and stronger AI research workflows.
xAI also said its safety evaluation reflects the model’s expanded capabilities, including what it described as its broadest pre-deployment capability and safeguard calibration testing to date, along with extensive post-deployment and third-party testing.
Benchmark results show mixed lead positions
Based on the public figures cited in the release, Grok 4.6 matches or exceeds GPT-5.6 Sol Max on some benchmarks, but not across the board. The scores listed in the article were:
- CursorBench v3.2: 69.9%
- GDPVal-AA v2: 1753
- AA-Briefcase: 1577
- APEX-Agents: 57.5%
- Terminal-Bench v3.0: 26%
- DeepSWE v1.1: 65.9%
xAI’s comparison said Grok 4.6 matched or beat GPT-5.6 Sol Max on CursorBench v3.2, GDPVal-AA v2, AA-Briefcase and APEX-Agents, while still trailing on Terminal-Bench v3.0 and DeepSWE v1.1. The article also said Fable 5 Max kept a narrow lead on most benchmarks.
API access and pricing
Beyond Cursor and Grok Build, xAI said Grok 4.6 is also available through its API and through partners including OpenRouter, Vercel and Cloudflare.
Pricing starts at $2 per million input tokens and $6 per million output tokens. xAI also offers a faster version priced at twice the standard model. The original post additionally points users to API key creation, documentation for integration, and a free trial path in Grok Build.