xAI ships Grok 4.6, built for longer agent runs

xAI released Grok 4.6 on August 12, reusing the same 1.5-trillion-parameter foundation as Grok 4.5 rather than the 2-trillion-parameter model Elon Musk had teased in July, with the gains coming from supervised fine-tuning and reinforcement learning instead of added scale. xAI’s own numbers put it level with GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61, ahead on CursorBench 3.2 (69.9%) and DeepSWE 1.1 (65.9%), and behind on Terminal-Bench v3.0 (26%). The API bills $2 per million tokens in, $6 per million out, roughly doubling for the faster variant.

The model is live now through the API, Cursor, Grok Build, and partners including OpenRouter, Vercel, and Cloudflare, with xAI offering 2x included usage in Grok Build and Cursor for the first week.

What it means for you

Treat “reused the same foundation, tuned harder” as the real headline: this is not a bigger or more capable Grok in the way the original 2-trillion-parameter teaser implied, it’s the existing model getting meaningfully better at coding and multi-step agent tasks specifically. If your agent workloads lean on tools like Cursor already, Grok 4.6’s CursorBench and DeepSWE numbers make it worth an A/B test against whatever you’re routing to now, the original 2T training teaser turning into a same-size, better-tuned release rather than a bigger one is worth remembering next time a lab teases a spec bump, and it pairs with Nvidia’s Nemotron 3.5 Lightning as another reminder to route tasks to the model that actually fits the job rather than defaulting to a single provider.