SpaceXAI launched Grok 4.6, a language model with 500,000-token context and a new reasoning level called xhigh. The version is a post-training update of Grok 4.5, not a larger base model, and is already available via the xAI API, on Cursor, and on Grok Build.
On the Artificial Analysis Intelligence index, the model scores 61 points, five more than Grok 4.5 and tied with GPT-5.6 Sol Max. SpaceXAI said the model underwent an additional round of training, with model-generated data, regenerated supervised fine-tuning trajectories, and reinforcement learning in agent environments. The company reported, based on internal tests, that the model checks its own work on long tasks, without independent measurement.
Benchmark performance
In coding tests, Grok 4.6 lagged behind competitors. On DeepSWE v1.1, it scored 65.9%, versus 73% for GPT-5.6 Sol Max. On Terminal-Bench v3.0, it reached 26%, almost double Grok 4.5, but still last among the four listed models. In other tests, the model scored 69.9% on CursorBench v3.2, 61.3% on FrontierCode v1.1 Extended, and 57.5% on APEX-Agents.
The results for GDPval-AA v2 and AA-Briefcase, in which the model appears ahead, are within the disclosed confidence margins, meaning they are statistical ties. The comparison table does not include Claude Opus 5, from Anthropic, which leads the index. The company did not disclose the parameter count of Grok 4.6.
Availability and pricing
Grok 4.6 is available as grok-4.6 on the xAI API, is the default model in Grok Build, and arrives on Cursor across all plans. Access is also possible via OpenRouter, Vercel, and Cloudflare. There is no open-weights version and no option for self-hosting.
The company announced prices of US$2 per million input tokens, US$0.50 for cached input, and US$6 for output on prompts with fewer than 200,000 tokens. Above that limit, the values rise to US$4, US$1, and US$12, respectively. A faster variant with a doubled price is mentioned on the launch page, but its identifier was not disclosed.
The model accepts text and image as input, generates text only, and has a knowledge cutoff of February 1, 2026. SpaceXAI recommends using prompt_cache_key or the x-grok-conv-id header to avoid full input charges for requests with cache.
