Grok 4.6 pricing revealed — DeepSeek 4 Pro undercuts it further

What one price sheet reveals

xAI's Grok 4.6 has come in at $2 per million input tokens and $6 for output. Compared to Claude Opus 5 ($5 input / $25 output) and GPT-5.6 Sol ($5 input / $30 output) on the same table, that's anywhere from half to a fifth of the price. Claude Sonnet 5 matches Grok 4.6's $2 input price but comes in higher on output at $10. Chubby, who runs an AI news account, shared the table and noted that "price is becoming more important than performance." They added that DeepSeek 4 Pro GA undercut even this price.

Performance is already top-tier — the fight now is over price

Grok 4.6 was already ahead in the performance debate. In the Artificial Analysis Intelligence Index released on August 12, Grok 4.6 scored 61, putting it in the top tier alongside GPT-5.6 Sol and up 5 points from its predecessor, Grok 4.5. On GDPval-AA v2, a practical-task benchmark, it posted an Elo of 1753, second only to Claude Opus 5 (1849), and it's reportedly already been integrated and tested in the coding tool Cursor. The key point of this price table is that Grok 4.6 cut its price without falling behind on performance.

Price competition isn't new in the AI model market. Running large language models costs money for every token processed, so if two models give the same answer, choosing the cheaper one is the rational move. The catch has been the assumption that "cheaper models perform worse" — and Grok 4.6 matching GPT-5.6-level intelligence while only lowering its price is shaking that assumption. On top of that, China's DeepSeek reportedly launched its latest model, DeepSeek 4 Pro GA, at a price even lower than Grok 4.6's, pushing the floor of price competition down once again.

So what changes now

Until now, model selection came down to a single benchmark score. Now it's shifting toward weighing "cost per point of performance." For developers and companies running large volumes of API calls, a difference of a few dollars per token can reshape an entire monthly bill. With Grok 4.6 delivering GPT-5.6/Claude-level performance at less than half the price, and DeepSeek pricing even lower still, model selection is moving away from "which model is smartest" and toward "which model can do the same job for the least money."