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

SpaceX Corp. unveiled Grok 4.7, its most advanced large language model yet, boasting enhanced long-horizon processing and safety upgrades. Originally developed by the Elon Musk-founded xAI Corp., xAI merged with xAI Inc. last year and was later absorbed into SpaceX, which now possesses the Grok model series and several AI data centers.

Grok 4.7 was assessed using CursorBench 4.0, a benchmark developed by Cursor, a recent SpaceX acquisition. Outperforming GPT-5.6 Sol and Fable 5.1, Grok 4.7 completed the challenges for an average cost of $4.69 per task, outpacing both models. The model was also tested against more commonly used benchmarks, outperforming Fable 5.1 on the Harvey Legal Agent Benchmark and EEBench, which focus on legal and chip design tasks, respectively; however, it trailed behind GPT-6 Astra on the latter.

SpaceX attributes its improved performance to a new base model and optimized reinforcement learning workflow, which involved tougher training tasks and extended durations. The LLM is designed to work with the Grok Bot harness, enabling parallel processing of complex tasks and accuracy verification among AI agents. SpaceX has fortified Grok 4.7 with new safeguards, setting records on LatchBio and HackerBench for blocking malicious biology research and cybersecurity requests.

Pricing starts at $2 per million input tokens and $6 per million output tokens, with an alternative version offering twice the speed for double the cost. Grok 4.7 launches shortly after SpaceX introduced a text-to-speech model, Grok Voice Transcribe 2.0, which boasts double the accuracy of its predecessor at half the price.