SpaceXAI Releases Grok 4.7

SpaceXAI has officially released Grok 4.7, now available across Cursor, Grok Build, and the Grok API.

Grok 4.7 introduces a larger base model and focuses reinforcement learning on long-duration tasks. The company reports improved detail in generative tasks, such as the urban game demo.

According to Choi citing Artificial Analysis, Grok 4.7 scored 46 points, a 2-point increase over version 4.6. It remains behind leaders like GPT-6 Astra and Fable 5.1, which scored 53 points.

Output tokens per task increased by 125% to 81,000, significantly higher than competitors. However, generation time improved, dropping from 9.5 minutes to 7.1 minutes.

Grok 4.7 showed strong gains in the AA-Briefcase benchmark, particularly in analysis quality. However, performance in output presentation quality saw a slight decline.

Coding agent performance rose from 47 to 56 points, surpassing GPT-5.6 Sol. Notable improvements were seen across DeepSWE and Terminal-Bench metrics.

Performance in long-context reasoning and automation benchmarks remained stagnant or declined slightly. Hallucination rates decreased, though overall accuracy remained largely unchanged.

Grok 4.7 maintains the pricing of its predecessor, offering a cost-effective alternative to GPT-5.6 Sol and Opus 5. It supports a 500k token context window.

CursorBench data suggests Grok 4.7 offers better cost-efficiency than Fable 5.1 and GPT-5.6 Sol in mid-range performance tasks. However, Fable 5.1 still leads in peak performance settings.

While Grok 4.7 provides valuable improvements in coding and cost-efficiency, its overall intelligence gains are limited. It is not yet a definitive rival to the absolute top-tier models.

Official release details are available at the x.ai news portal.