Grok 4.6 Released: xAI's Cursor Acquisition Ushers in a Three-Way AI Race

A Three-Way Race: The AI Competition Gets a Third Contender

For the past two years, the industry consensus was that only two players dominated the AI large model space—OpenAI and Anthropic. But in just a matter of days, a new formidable competitor has emerged that deserves serious attention. According to tech blogger Matthew Berman's analysis, Grok (xAI's model) once made a splashy entrance but quickly faded into silence with little action for a long time. Now, all of that has fundamentally changed.

The turning point came when xAI acquired Cursor, completing the critical missing piece it needed to become a dominant player. Just recently, xAI officially released its latest model, Grok 4.6—a model that is quite unique across multiple dimensions.

It's worth noting that Grok 4.6 is not an entirely new model trained from scratch, but rather a ".x" iterative upgrade built on Grok 4.5. However, it represents a massive improvement over its predecessor, particularly in coding and knowledge work capabilities—which is precisely xAI's current core strategic focus.

Grok 4.6 Benchmarks: Comprehensive Improvements with Outstanding Cost-Efficiency

Looking at benchmark results, Grok 4.6's performance is quite impressive. The previous Grok 4.5 was already the default sub-agent model in Cursor, known for its strong capabilities and speed. Grok 4.6 improves on nearly every front and is even faster.

Specifically:

• GDP Val (OpenAI's benchmark measuring real-world knowledge work task effectiveness): Grok 4.6 high takes first place, surpassing competitors
• Cursor Bench: Surprisingly didn't take first, but is essentially tied with Fable 5 Max
• DeepSuite (considered by many to be the benchmark closest to real coding experience): Grok 4.6 scores 65.9, ranking third, behind GPT 5.6 Sol Max (73) and Fable 5 (70)
• Terminal Bench: Jumped from 15% in the previous generation to 26%
• Harvey Lab (legal scenarios): Dominates with a 15.8% score

Cost Per Task Is the Key Metric

Matthew Berman specifically emphasizes that looking at quality scores alone isn't enough—cost matters equally. On Artificial Analysis's intelligence index leaderboard, Grok 4.6 high leaped from 4.5's position to fourth place, tied with GPT 5.6 Sol, behind only Claude Opus 5, Fable 5, and GPT 5.6 Sol.

But what's even more noteworthy is the "cost per task" dimension. This depends on two factors: how many tokens a model consumes and the price per token. Grok 4.5 high's cost per task was approximately $0.36 with an intelligence index of about 0.55; while Grok 4.6's intelligence index improved to 0.60, its cost also rose to approximately $0.83. Although it's more expensive, compared to GPT 5.6 Sol Max—which has a similar intelligence level—Grok 4.6 is still cheaper.

Hands-On Comparison: Grok 4.6's Design Capabilities Still Have Gaps

Berman used the same prompt to have three top models generate a visually appealing single profile card. The results were quite dramatic:

• Fable 5: Performed terribly, generating an image described as "the worst SVG drawing software output"
• GPT 5.6 Sol: Generated a beautiful portrait image with clean design and strong visual appeal, clearly the winner
• Grok 4.6: Decent but not outstanding—while it also generated a portrait image, it had issues like a cropped follow button, no bottom margin, and awkward text spacing

Although design benchmarks are highly subjective, from hands-on testing, GPT 5.6 clearly leads in this area, while Grok 4.6 sits in the middle tier.

Grok 4.6 Pricing and Availability: Positioned as a Workhorse Model

Grok 4.6 is now live and available through Cursor, Grok build, the API, and channels like OpenRouter, Vercel, and Cloudflare. Pricing is set at $2 per million input tokens and $6 per million output tokens—quite competitive.

For comparison, GPT 5.6 Sol costs several times more, and Fable is even pricier. This is why Berman positions Grok as a "workhorse model." There's also a Fast variant that costs double but runs faster, and Grok build and Cursor users get double usage credits during the first week.

xAI's Cursor Acquisition: The Perfect Marriage of Data and Compute

To understand xAI's current position of strength, we need to look back at the story behind this acquisition. In April of this year, xAI acquired Cursor. Cursor was one of the first companies to deeply integrate AI into an IDE, pioneering an entirely new approach to AI-assisted coding. However, after Claude Code and Codex launched, Cursor fell behind as developers flocked to other products.

This acquisition was critical because the two sides formed a perfect complement: Cursor possessed massive, highly valuable coding data but lacked the data centers and GPUs to train its own models; meanwhile, xAI built 200,000 GPUs in just 122 days at an astonishing pace but lacked a popular, capable model, leaving much of that GPU capacity idle. When data and compute merged, powerful coding models like Grok 4.5 and now 4.6 were born.

Recursive Self-Improvement: Training Grok with Grok

Interestingly, xAI explicitly stated in their blog post that they used Grok 4.5 to train Grok 4.6—using the previous model to regenerate SFT training trajectories covering reasoning, agentic frameworks, and domains like STEM, software engineering, and knowledge work. While this isn't yet fully closed-loop self-improving AI, current leading labs are indeed all using their models to train next-generation models, and Grok is no exception.

Grokbot: An AI Coding Agent for the Masses

Grok 4.6 also powers Cursor's newest product, Grokbot. Unlike Cursor, which targets developers, Grokbot is explicitly designed for a broader, less technical user base. It strips away model selection—users can't even tell which model they're using—and displays no code at all, showing only text and outputs like presentations and Word documents.

In the interface, each conversation thread is an agent, and each agent has its own little mascot. This "powerful yet extremely simple to use" design philosophy reflects xAI's strategic intent to reach a broader market—after all, the entire knowledge work market is fundamentally built on code.

AI Competitive Landscape: Warning Signs for Anthropic

One intriguing detail: Anthropic has struck a deal with xAI and is purchasing compute capacity from them. Berman describes this as "like selling weapons to your enemy," and given Musk's previous negative comments about Anthropic, this deal must be quite satisfying for him.

But looking at it from the other side, if Grok 4.6 continues to attract demand, xAI will very likely redirect all GPU capacity toward Cursor and Grok itself once the contract expires. For Anthropic, this is a warning signal worth heeding.

And xAI isn't slowing down. Musk has already tweeted that Grok 4.7 is "significantly better than 4.6," expected to be ready in three to four weeks, and will incorporate substantial supplementary training data from companies under his umbrella like SpaceX.

Conclusion: Betting on Coding Is Betting on the Future of AI

From Grok 4.6's release, an industry pattern becomes crystal clear: betting on coding is betting on the future. This playbook of "deeply investing in coding use cases → building a data flywheel → training the next-generation model" was first validated by Anthropic with Claude Code, then replicated and matched by OpenAI through Codex, and now successfully adopted by xAI following its Cursor acquisition.

The AI race has officially entered a three-way era. While xAI still ranks third in overall capability, its accelerating momentum cannot be ignored. For users, fiercer competition means better and cheaper models—which is undoubtedly a good thing.