xAI Releases Grok 4.7: Parameter Scale Increased to 2.1 Trillion, Still Trails Claude in Benchmark Tests

动察 Beating AI News Flash: xAI officially released Grok 4.7 on Monday, claiming a "significant improvement" over Grok 4.6 at the same price and speed. Grok 4.7 had been delayed at least 5 times and is now available on the Grok app, Cursor, Grok Build, and the xAI API.

Grok 4.7 has a parameter scale of 2.1 trillion, a 40% increase over Grok 4.6's 1.5 trillion, with input and output pricing at $2 and $6 per million tokens, respectively. xAI also introduced supplementary training data from SpaceX, including Starlink satellite telemetry data, manufacturing records, and engineering failure logs, aiming to improve the model's reasoning capabilities for hardware and physical systems.

In terms of benchmarks, Grok 4.7 ranked second in both GDPval and AA-Briefcase, scoring 1695 and 1657, respectively, both trailing Claude Fable 5.1's 1735 and 1678; in the electrical engineering benchmark EEBench, it ranked second, behind only GPT-6 Astra. CursorBench 4.0 shows that Grok 4.7's task cost is higher than GPT-6 Astra and Claude Sonnet 5, while its overall performance still lags behind Fable 5.1.

Musk previously said that Grok 4.7 is roughly comparable to Claude Opus 5.0, but has not yet reached the level of the newer version Opus 5.1, and revealed that Grok 4.8, 4.9, and Grok 5 are under development.