xAI Unveils Grok 4.7, Betting on Long Tasks and Safety

On September 21, 2026, xAI introduced Grok 4.7, which it calls its most capable model for coding and knowledge work. The release is offered at the same price and speed as the earlier Grok 4.6, positioned as competitive within its class.

What changed

Grok 4.7 is built on a new, larger base model. It was trained with a longer reinforcement-learning run on a harder mix of tasks, weighted toward problems that take many hours to finish. The developers stress that the model verifies its own work more carefully and handles long context more reliably. It was also trained to natively understand the Grok Bot harness, improving conversational scenarios and general knowledge work.

On the benchmarks cited in the announcement, Grok 4.7 outpaces its predecessor: 46.3% versus 40.4% on CursorBench 4.0, 64% versus 53% on the engineering-focused EEBench, and 19.6% versus 15.8% on the Harvey legal test. On several metrics it is comparable to rivals such as GPT and Fable, and trails them on others.

Safety

xAI puts particular emphasis on a new safeguard stack. According to the company, this is the strongest model it has tested for refusals and jailbreak resistance. In dual-use domains — cybersecurity and biology — it leads on the balance between usefulness for benign tasks and refusals of dangerous ones: 62.4% on the LatchBio biosafety benchmark and only 3.3% of risky prompts let through on its own HackerBench v0.3 test. Select cybersecurity partners have been granted invite-only access to its red-team capabilities for defensive research.

Price and availability

The model is already available in Cursor and Grok Build, as well as through the API, third-party coding harnesses, model routers, and cloud platforms. Base pricing is $2 per million input tokens and $6 per million output tokens. There is also a "fast" variant with twice the output speed at twice the price.

Source: Hacker News