Grok 4.7 Keeps AI Prices Low — But There's a Catch

SpaceXAI launched Grok 4.7 on September 21, a new AI model aimed at coding and professional knowledge work. It's a bigger model than Grok 4.6 and was trained longer on hard, multi-step tasks — the kind that can take hours. On company-run coding tests, it beat its predecessor, roughly doubling its score on one challenging test. That matters if you're paying developers to write and check code.

Here's why the price tag isn't the whole story. Grok 4.7 costs the same as before: about $2 per million words of input and $6 per million words of output (tokens are the small text chunks AI reads and writes). But independent testers at Artificial Analysis found Grok 4.7 used around 81,000 tokens per task at its highest reasoning setting — far more than competing models. Think of it like a taxi that charges less per mile but takes a much longer route. For businesses, what counts is the cost of one finished job, not the advertised rate.

SpaceXAI also added new safety features. The company says Grok 4.7 is harder to trick into saying dangerous things, while refusing fewer harmless requests. It reports a 62.4% score on a biosafety test and says only 3.3% of risky test prompts slipped past its filters. Selected cybersecurity partners are getting early invite-only access to stress-test it. These numbers come from SpaceXAI itself, so outside checks will matter.

Practically, Grok 4.7 is built for long, repetitive work: digging through code, running tools over and over, and handling lengthy documents or presentations. It's available through Cursor, the Grok API, and major cloud platforms. The honest catch is predictability — results shift depending on settings, and heavy token use can quietly raise real-world costs.