Elon Musk's SpaceXAI released Grok 4.7 on Monday at $2 per million input tokens, though its own benchmarks put Anthropic's Claude Fable 5.1 ahead on most tests.
Key Points:
- SpaceXAI launched Grok 4.7 at the same price as Grok 4.6, with output tokens costing $6 per million.
- Claude Fable 5.1 kept the lead on GDPval, CursorBench and Terminal-Bench in the company's own comparison.
- Grok 4.7 won on electrical engineering and legal work after months of shifting launch dates.
Grok 4.7 Rollout
The company announced the model on Monday as its most capable yet for coding and knowledge work, and it is live in Cursor, Grok Build, the Grok API and third-party coding harnesses. A faster variant costs twice as much. Output tokens run $6 per million, unchanged from Grok 4.6, which shipped just over a month ago.
SpaceXAI said Grok 4.7 runs on a new, larger base model trained with a longer reinforcement learning run weighted toward problems that take many hours to finish, and it ships with a rebuilt safeguard stack. Musk has separately claimed the model carries 2.1 trillion parameters, up from 1.5 trillion, with supplemental training drawn from SpaceX engineering data that the launch post never mentions.
Also Read: Polymarket Bug Let Thieves Walk Into Accounts Without A Password
Claude Fable 5.1 Lead
On GDPval, which scores AI on work done by lawyers, nurses and financial analysts, Grok 4.7 posted an Elo rating of 1,695 against 1,735 for Claude Fable 5.1. It finished behind again on AA Briefcase, a test of multi-hour office tasks. On Terminal-Bench the gap widened to almost 20 points.
The newer model did win in places, beating Fable 5.1 on an electrical engineering test, 64% to 56.4%, and tripling its rival's score on Harvey's legal agent benchmark for lawyers. Grok 4.7 improved on Grok 4.6 across every test the company published, nearly doubling its predecessor's score on terminal work.
Each of those figures comes from SpaceXAI itself. CursorBench, the coding test the company led with and lost 51.8% to 46.3%, is built by Cursor, the editor SpaceX finished buying last month. What Grok 4.7 does hold is price, at a fifth of what Fable 5.1 charges for input tokens.
Elon Musk Delays
The launch closed out a summer of missed targets, with Musk pushing the date back repeatedly after saying in late July that the model was four weeks away. He narrowed the window to 10 days on Sept. 1, then conceded on Sept. 11 that the model needed a few more days of training.
Days before shipping, he lowered expectations again, pegging Grok 4.7 at roughly Claude Opus 5.0 rather than the newer Opus 5.1, with multimodal work still unfinished. He has since called Grok 4.8 a meaningful step up. Grok 5, he suggested, could lead the field outright, though neither model has a release date.
Read Next: AI Is Not Ready For Its Own Weapons, Warns The Hacker Who Broke Into OpenAI
