SpaceXAI released Grok 4.7 on 2026-09-21, about six weeks after Grok 4.6. The company calls it “our most capable model for coding and knowledge work,” saying it “works longer on difficult tasks” and “checks its own work more carefully.” Unlike the 4.6 update, which reused the 4.5 base, Grok 4.7 is built on “a new, larger base model,” trained with “a longer reinforcement learning run on a harder mix of tasks, weighted toward problems that take many hours to complete,” and natively trained for the Grok Bot interface.
Pricing is unchanged from Grok 4.6 at 2 dollars per million input tokens and 6 dollars per million output, with a fast variant at double the price for roughly twice the output speed. The model shipped the same day in Cursor, Grok Build, the Grok API and third-party coding platforms and routers. SpaceXAI’s marketing line is “twice as fast, at half the price of comparable models.”
Published results include CursorBench 4.0 at 46.3 percent against 40.4 for Grok 4.6, DeepSWE v1.1 at 71.0 percent (up from 65.9 for 4.6), 64.0 percent on an electrical-engineering benchmark (EEBench), 19.6 percent on Harvey’s legal benchmark, and 62.4 percent on a LatchBio biology benchmark. On the safety side, SpaceXAI reports that only 3.3 percent of risky prompts got through on its HackerBench cybersecurity evaluation.
Why it matters: Grok 4.7 arrived one day before Claude Opus 5.5 and GPT-6 Sol and Luna, making the week of 2026-09-21 a three-way price-and-agent-performance contest in the tier just below each lab’s flagship. What it does not show: all numbers are SpaceXAI’s own, several benchmarks are vendor-specific or partner-run, and the “half the price of comparable models” framing depends on which models the company chose to compare against.