xAI, the company founded by Elon Musk, has released Grok 4.7, its most capable model to date for coding and knowledge work, at $2 per million input tokens and $6 per million output tokens. The price is the central fact of the launch: it places the model closer to Chinese offerings than to Western frontier systems. For companies that pay for inference at scale, that difference matters more than a headline benchmark score.
What is inside Grok 4.7
According to the company, the model is built on a larger base model than its predecessor and was trained with longer reinforcement learning. A separate design goal was better verification of its own output, which xAI presents as a way to reduce errors in multi-step tasks. The model is available through the Grok API, Cursor and Grok Build, so buyers can reach it both directly and inside existing coding tools. No parameter count, training compute figure or release date beyond the launch itself was disclosed in the source.
The pricing of $2 per million input tokens and $6 per million output tokens is what separates Grok 4.7 from the Western frontier. Claude Fable 5.1 and GPT-6, the two models it is measured against, are positioned as premium systems, and the source notes that the low rate is probably set for a good reason. In practice, a team running large volumes of code review, document processing or agent loops pays per token on every call, so the input price drives the monthly bill more than any single capability test. The trade-off is explicit: lower cost per call against weaker results on the tasks that are hardest to automate.
What this means for business
On the independent Artificial Analysis Intelligence Index (v4.3.2), which combines ten benchmarks, Grok 4.7 scores 46 and lands mid-pack. Claude Fable 5.1 and GPT-6 lead with 53 each. The seven-point gap is the first signal for anyone choosing a default model: for summarization, drafting and routine knowledge work, the difference may not be visible in day-to-day output, while the cost difference is visible in every invoice. A small company with limited inference budget can put Grok 4.7 into production and keep the premium models for the narrow set of tasks where quality is measurable.
The picture changes in agentic coding, where a model has to act inside a terminal rather than answer a question. On Terminal-Bench 4.0, Grok 4.7 reaches just 26 percent, against 60 percent for GPT-6 Astra and 55 percent for Claude Fable 5.1. Even the cheaper DeepSeek V4.1 Flash edges past it at 27 percent. That result is the practical limit of the launch: for teams building autonomous coding agents, the low token price does not compensate for a pass rate less than half that of the leaders, because a failed run costs engineering time and a repeated attempt costs tokens again.
Before adopting the model, buyers should separate the two claims. The price and the availability through the Grok API, Cursor and Grok Build are confirmed facts of the launch. The claim about better self-verification comes from xAI itself and is not backed by an independent measurement in the source, so it needs testing on the company's own tasks. The same applies to the mid-pack index score: ten benchmarks averaged into one number say nothing about performance on a specific domain, and the agentic coding result suggests where the weakness sits. Questions worth putting to a vendor are which benchmark version was used, whether the evaluation ran on the same harness as the competing models, and what the pass rate looks like on the buyer's own repository.
The trend will be confirmed or refuted by how the price gap moves. If Claude and GPT-6 hold their rates while xAI keeps Grok 4.7 at $2 and $6 per million tokens, cheap inference becomes a permanent second tier for routine work, and the premium tier stays reserved for agentic coding and other tasks where the pass rate decides the outcome. If the leaders cut prices in response, the gap that matters for business closes, and the benchmark spread becomes the only remaining argument.
