SpaceX launches Grok 4.7 with long-horizon processing, safety upgrades
SiliconANGLE Maria Deutscher ● Covered by 3 sources
SpaceX shipped Grok 4.7, its strongest model yet, with faster long-task handling and new safety guardrails. It also undercut some rivals on cost in one benchmark, while a faster tier costs twice as much.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
SpaceX has launched Grok 4.7, the latest and most capable model in its Grok line. The company says the new version came out of a changed training recipe: a new base model, tougher reinforcement learning, and longer runs on harder tasks than before.
The model’s performance story is tied to benchmarks, and SpaceX is leaning hard on them. On CursorBench 4.0, Grok 4.7 finished tasks at an average cost of $4.69 each, ahead of GPT-5.6 Sol and Fable 5.1. SpaceX says it tested hardware-heavy versions of those rivals that favor quality over thrift, which makes the comparison more interesting than a simple scoreboard.
It also did well on the Harvey Legal Agent Benchmark and EEBench, where the company says it beat Fable 5.1 on both. EEBench covers chip design tasks, and on that test Grok 4.7 still trailed GPT-6 Astra, OpenAI Group PBC’s latest model. So this is not a clean sweep. But it is a pretty strong debut.
The model is built to work with Grok Bot, SpaceX’s setup for splitting complex jobs across multiple AI agents. Those agents can run tasks in parallel and check each other’s output, which should help with longer, messier workflows. SpaceX is pitching that as long-horizon processing, and the phrase fits: this model is less about a flashy one-shot answer and more about staying organized when the work sprawls.
Safety is part of the pitch too. SpaceX says Grok 4.7 set records on LatchBio and HackerBench, which measure how well a model refuses malicious biology and cybersecurity requests. Pricing starts at $2 per million input tokens and $6 per million output tokens, with a version for latency-sensitive jobs that is twice as fast and twice as expensive. The timing is notable: this comes less than a week after SpaceX’s previous release, a text-to-speech model called Grok Voice Transcribe 2.0.
My take — AI-written commentary, not fact-checked reporting
The real tell here is not the benchmark parade, it’s the pricing split. Open models keep getting sold as cheap and nimble while closed systems collect premiums for speed, safety, and agent plumbing. SpaceX is betting customers will pay for the plumbing, which is exactly how these products start to look less like models and more like toll roads.
Read more about this at: SiliconANGLE
