Introducing Grok 4.6. It delivers frontier intelligence and is a significant improvement over Grok 4.5 at the same price. Grok 4.6 is faster than comparable models and can handle much more challenging tasks than Grok 4.5. It's half the price of other frontier models at $2/M input and $6/M output tokens. Grok 4.6 is available today in Grok Build, Cursor, Grok Bot, and the API. We’re including 2x usage inside Cursor and Grok Build for the first week. https://x.ai/news/grok-4-6
The price drop is the biggest thing that jumps out to me here considering the benchmarks are the same or better than Grok 4.5 (and many other frontier models). I also wonder if we’ll see an early spike in usage since I assume this will quickly become a commonly used model within Cursor and OpenRouter
Half the price of other frontier models is a bold claim given how tight margins already are on inference. I'd like to know how they're pulling that off at the same output quality, efficiency gains on serving or something in the training run itself.
The new Grok 4.6 pricing is a massive shake-up for the API market. At $2/M for input, it’s effectively undercutting GPT-5.6 Sol and Claude Opus 5 by nearly 60-80%, especially on output costs. For developers running high-volume agents or heavy reasoning workflows, this kind of price-to-performance ratio ($0.84/task vs. ~$3/task for competitors) makes it a no-brainer to test. Just gotta watch out for that price doubling if you hit the 200k context cliff
The lower token price is a meaningful change, especially for teams that have been hesitant to test agent workflows at real volume. The part we would still measure is cost per successful task, including retries, tool calls, and review time. Cheap tokens help, but a small workflow that works reliably is what makes automation feel useful.
An interesting advancement. The competition among frontier models is no longer just about benchmarks; efficiency, inference cost, and the ability to operate at scale are becoming just as important. In the end, the balance between performance, cost, and integration will be critical for real-world AI adoption in enterprises.
The pricing is interesting, but the bigger shift is model routing becoming normal infrastructure. If Grok 4.6 can deliver stronger performance at lower cost, teams get more flexibility to match model capability to the task instead of defaulting to one expensive model for everything. The real advantage will come from choosing the right model per workflow, then measuring quality, latency and cost together.
The interesting claim isn’t the price; it’s that longer trajectories show more self-testing. The hard part is independence: when the same model family writes the plan, executes it, and judges the result, correlated blind spots can survive every layer. Has the team measured how often an independent verifier rejects a self-approved result as trajectory length grows?
SpaceXAI This chart is less a coronation than a commodity-market diagram. Grok wins GDPVal, loses CursorBench, DeepSWE, FrontierCode, APEX-Agents, Terminal-Bench and APEX-SWE, then competes aggressively on $2/$6 pricing. That is a healthy frontier model market, but it is the opposite of durable model supremacy: several suppliers are close enough that price, routing and workload determine the winner. If customers can swap the intelligence underneath their agent without rebuilding the system, where exactly is Musk’s moat: the model, or merely this month’s position on the router?
The pricing is what really stands out here. If Grok 4.6 can deliver frontier-level performance at $2/M input and $6/M output, the competitive pressure on inference economics is getting serious.