Grok 4.7

SpaceXAI has unveiled a new model named Grok 4.7, which is touted as the most powerful model for coding and knowledge work. This model is touted as being twice as fast as its predecessor, Grok 4.6, at half the cost. Grok 4.7 is designed to work more extensively on complex tasks, scrutinize its output more meticulously, and boasts the best safety measures yet.

The model is priced at the same affordability as Grok 4.6 and operates at the same speed. Grok 4.7 has been evaluated on CursorBench 4.0, a test that focuses on lengthy coding tasks, and has emerged as a frontrunner in terms of price-performance.

Grok 4.7 is trained on a larger base model and underwent a more rigorous reinforcement learning phase, focusing on problems that can take several hours to solve. The model has been enhanced to verify its own work better and manage larger context. Additionally, it has been fine-tuned to understand the Grok Bot harness more naturally, making it more adept at conversational tasks and general knowledge work.

Grok 4.7 excels in generating documents and presentations. It outperforms Grok 4.6 in GDPval and AA Briefcase benchmarks, where AI is tasked with professional duties like those of lawyers, nurses, and financial analysts. The model also demonstrates superior performance compared to other frontier models in these areas.

Grok 4.7 has been built with an entirely new safeguard stack, making it the strongest model tested in regards to refusal and jailbreak resistance. In dual-use fields like cybersecurity and biological work, it excels in utility for benign tasks and safe refusal on hazardous ones. It has outperformed LatchBio's biosafety benchmark with a score of 62.4%.

The model strikes a balance between robust cyber defense capabilities and minimal refusal rates for legitimate use. It performs exceptionally well on HackerBench v0.3, a benchmark designed for risky and malicious cyber tasks, allowing only 3.3% of risky dual-use prompts to pass while rarely obstructing legitimate security work. SpaceXAI has also introduced invite-only access to Grok 4.7's red-team capabilities for cybersecurity partners to aid in defense research.

Grok 4.7 is currently available in Cursor and Grok Build platforms and can also be accessed via the Grok API, third-party coding harnesses, and model routers and cloud platforms. Pricing starts at $2 per million input tokens and $6 per million output tokens. A faster variant of the model is also available, with double the output speed at double the price.