SpaceX launches Grok 4.7 with long-horizon processing, safety upgrades

SpaceX Corp. today introduced Grok 4.7, its most capable large language model to date.

The Grok algorithm series was originally created by an Elon Musk-founded startup called xAI Corp. Musk merged the company with xAI Inc. last year. The combined organization was subsequently folded into SpaceX, which gained not only the Grok model series but also multiple artificial intelligence data centers.

The company evaluated Grok 4.7 using a benchmark called CursorBench 4.0. It was developed by Cursor, another of SpaceX’s recent acquisitions. Grok 4.7 completed the challenges in the benchmark for an average cost of $4.69 per task, which put it ahead of GPT-5.6 Sol and Fable 5.1. SpaceX evaluated hardware-intensive versions of the latter models that prioritize output quality over cost-efficiency.

The company also tested Grok 4.7 using more widely-used benchmarks. The model handily outperformed Fable 5.1 on the Harvey Legal Agent Benchmark and EEBench, which contain legal and chip design tasks, respectively. However, Grok 4.7’s score on the latter benchmark fell behind GPT-6 Astra, OpenAI Group PBC’s latest LLM.

SpaceX credits Grok 4.7’s performance to a new base model. LLM training runs comprise multiple phases that each improve a different aspect of the algorithm being developed. A base model is the initial version of an LLM that emerges after the first phase of a training run. AI providers often reuse base models across LLM releases.

According to SpaceX, its engineers also enhanced Grok 4.7’s reinforcement learning workflow. Reinforcement learning is an AI training method used to enhance base models’ reasoning capabilities. Compared with its predecessor, Grok 4.7 received tougher training tasks and worked on them for longer.

The company built the LLM to be compatible with its Grok Bot harness. It’s a collection of technical resources that enable Grok 4.7 to 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 has also equipped Grok 4.7 with new safeguards. The company says the model has set records on LatchBio and HackerBench, benchmarks that test LLMs’ ability to block malicious biology research and cybersecurity requests, respectively.

Grok 4.7 is available from $2 per million input tokens and $6 per million output tokens. For latency-sensitive  workloads, SpaceX is offering a version of the model that processes prompts twice as fast and costs twice as much.

The model is launching 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 and costs half as much. SpaceX says that Grok Voice Transcribe 2.0 also outperforms several of its competitors’ entries into the text-to-speech category.

Photo: Unsplash