Grok 4.7 Targets Long-Horizon Knowledge Work
Grok 4.7 is a new model announced by SpaceXAI on September 21, 2026, with a focus on coding and knowledge work. The company says it uses a larger base model and a longer reinforcement-learning run aimed at tasks that can take hours. It is also designed to manage longer context, verify its own work, and produce documents and presentations more effectively.
Pricing and Benchmarks Form the Main Pitch
API pricing starts at $2 per million input tokens and $6 per million output tokens, the same prices listed for Grok 4.6. A fast variant doubles output speed at twice the price. The company reports a 46.3% score for Grok 4.7 on CursorBench 4.0, compared with 40.4% for Grok 4.6, along with competitive results on DeepSWE, EEBench, and AA Briefcase. These tests measure different tasks, however, and some results use a high-effort mode, so no single score fully predicts real-world performance.
Safety Claims and the Open Questions
SpaceXAI also highlights a new safeguard stack. In the company’s disclosed testing, Grok 4.7 scored 62.4% on a biosafety benchmark and allowed only 3.3% of risky dual-use prompts through on HackerBench v0.3, while aiming to avoid blocking legitimate security work. The model is available in Cursor, Grok Build, the Grok API, third-party coding harnesses, and cloud platforms. The main limitation is that these performance and safety figures come from the publisher and the cited benchmarks; independent verification is not provided in the source. Actual results will also depend on the task, tools, and context supplied.