Grok 4.6: xAI Matches GPT-5.6 at Half the Price
The new flagship reaches GPT-5.6 Sol's benchmark level while costing far less. The clear focus is on agentic tasks and long-running workflows.
Illustration · AI-generated (AI IN LIFE)
At a glance
- Scores 61 on the Artificial Analysis Intelligence Index — level with GPT-5.6 Sol, one point behind Claude Fable 5
- Elo 1753 on the agentic benchmark GDPval-AA v2
- Pricing unchanged: $2 per million input tokens, $6 per million output tokens
- Same 1.5-trillion-parameter V9 architecture as Grok 4.5
- Available via API, Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare
xAI — now operating under the SpaceX umbrella — has released Grok 4.6. The model scores 61 on Artificial Analysis's Intelligence Index, level with OpenAI's GPT-5.6 Sol and just one point behind Anthropic's Claude Fable 5. That is a five-point jump over its predecessor, Grok 4.5.
What makes the jump notable is how it was achieved: not with a bigger model, but through a longer supplemental training run plus improved supervised fine-tuning and reinforcement learning. The architecture is unchanged — the same V9 base with roughly 1.5 trillion parameters as Grok 4.5.
The real emphasis is on agentic work. On GDPval-AA v2, a benchmark measuring real-world knowledge tasks, Grok 4.6 reaches an Elo rating of 1753, completing tasks in around 53 turns. xAI is positioning the model squarely against the frontier tier for long-running agents, coding, and multi-step workflows.
On pricing, xAI is attacking: $2 per million input tokens and $6 per million output tokens — unchanged from the predecessor and roughly half of what comparable frontier models cost. Grok 4.6 is available via the API and through partners including Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare, with doubled usage quotas during the first week.
One announcement from Elon Musk is drawing debate: future versions are to be trained on “the sum total of all SpaceX information.” The company has not specified which internal data that covers or whether employees can opt out — a point observers are watching critically.
FAQ
What is new in Grok 4.6?
Mainly its agentic performance: the model reaches frontier level even though the architecture is unchanged. The gains come from better training, not more parameters.
How much does Grok 4.6 cost?
$2 per million input tokens and $6 per million output tokens — about half the price of comparable frontier models, with doubled usage quotas in the first week.
Why is training on SpaceX data controversial?
Musk announced that future models will train on internal SpaceX information. It remains unclear which employee data is included and whether workers can opt out.
