SpaceX Corp. introduced Grok 4.7, its most capable large language model to date, and priced it from $2 per million input tokens and $6 per million output tokens. On CursorBench 4.0, a benchmark built by Cursor, another recent SpaceX acquisition, the model completed tasks at an average cost of $4.69 per task, ahead of GPT-5.6 Sol and Fable 5.1. The release matters because cost per completed task, not raw benchmark rank, is what determines whether an agent can be run in production at scale.
What the benchmarks show
The company tested Grok 4.7 on several widely used benchmarks as well. It outperformed Fable 5.1 on the Harvey Legal Agent Benchmark and EEBench, which contain legal and chip design tasks respectively. On EEBench, however, its score fell behind GPT-6 Astra, the latest LLM from OpenAI Group PBC. SpaceX compared its model against hardware-intensive versions of the competing systems, configurations that prioritize output quality over cost-efficiency. For latency-sensitive workloads, the company offers a variant that processes prompts twice as fast and costs twice as much, so buyers can trade price against response time within the same model family.
SpaceX attributes the performance to a new base model. LLM training runs consist of multiple phases, each improving a different aspect of the algorithm, and the base model is the initial version that emerges after the first phase. Providers often reuse base models across releases, which makes a fresh base a meaningful change rather than a routine update. Engineers also reworked the reinforcement learning workflow, the training method used to strengthen a base model's reasoning. Compared with its predecessor, Grok 4.7 received tougher training tasks and worked on them for longer, an approach the company describes as long-horizon processing.
The model is built to work with the Grok Bot harness, a collection of technical resources that lets Grok 4.7 split complex work among multiple AI agents. Those agents can perform several different tasks in parallel, which speeds up processing, and verify the accuracy of each other's output. SpaceX also equipped the model with new safeguards, saying it set records on LatchBio and HackerBench, benchmarks that test an LLM's ability to block malicious biology research and cybersecurity requests. The combination of parallel agents and refusal testing points to enterprise deployment, where an unchecked request is a liability rather than a feature.
The launch comes less than a week after SpaceX's previous model release. On Friday, the company introduced a text-to-speech model that offers double the accuracy of its predecessor at half the cost, and says Grok Voice Transcribe 2.0 outperforms several competitors in the category. The cadence reflects how the Grok line has been consolidated: the algorithm series began at xAI Corp., the startup founded by Elon Musk, which merged with xAI Inc. last year before the combined organization was folded into SpaceX. That structure gave SpaceX not only the model series but also multiple artificial intelligence data centers.
What this means for business
For companies adopting AI, the practical shift is that a capable model now carries a published per-task cost. At $4.69 per completed CursorBench task, teams can estimate the expense of an agent workflow before committing to it, and the $2 and $6 per million token rates make high-volume input processing the cheaper side of the bill. A small company can start with the standard version and pay only for what it runs, while a larger organization with latency-sensitive workloads may accept the doubled price for twice the speed. The parallel-agent harness matters most where several checks must happen at once, such as document review or code analysis.
Several things remain unverified. The benchmark results come from SpaceX itself, and independent reproductions are not yet available. The comparison against hardware-intensive versions of rival models means the cost advantage may not hold against their standard configurations, so buyers should ask which configuration was measured. The safety records on LatchBio and HackerBench describe refusal behavior on specific test sets, not a guarantee for every deployment. Before choosing, ask the vendor for per-task costs on your own workload, the latency of the faster tier, and how the agent harness handles conflicting outputs. The release does not by itself prove that Grok 4.7 is the best model for a given task.
The marker to watch is whether independent evaluations reproduce the CursorBench 4.0 cost figure and the EEBench gap against GPT-6 Astra within the coming weeks. If third parties confirm the per-task economics, per-task pricing becomes a standard way to compare models, and procurement teams gain a number they can put in a budget. If the results do not replicate, the advantage stays a vendor claim and buyers should keep testing candidates side by side.
