xAI unveils Grok 4.7: more power for coding and knowledge work at the same price as Grok 4.6

xAI announced on September 21 the launch of Grok 4.7, described as its most capable model to date for programming and knowledge work. According to the company, the model spends more time on hard tasks, checks its own work more carefully and arrives with what the firm calls its best-calibrated set of safeguards so far. One detail that matters for the market is that it is served at the same price and speed as its predecessor, Grok 4.6, which in practice means this is not a cost cut in absolute terms but rather more capability for the same outlay.

On the technical side, xAI says Grok 4.7 uses a larger model base than Grok 4.6 and was trained with a longer reinforcement learning run on a harder mix of tasks, weighted in particular toward problems that require many hours of sustained work. The company claims the model improves at verifying its own work and at handling longer contexts. It was also trained to natively understand the environment known as "Grok Bot", which according to xAI translates into better performance on conversational and general knowledge tasks.

In performance-versus-cost benchmarks, the company presents Grok 4.7 as sitting on the price-performance frontier within CursorBench 4.0, a test that emphasizes long-running programming tasks, comparing it with models from other providers such as Fable 5.1, Opus 5, GPT-5.6 Sol and Sonnet 5. In a more detailed table comparing specifically Grok 4.7 (xHigh variant), Grok 4.6 (High), GPT-5.6 Sol (Max) and Fable 5.1 (Max), prices per million tokens are $2 and $6 (input and output) for Grok 4.7 and Grok 4.6, against $4 and $20 for GPT-5.6 Sol, and $10 and $50 for Fable 5.1.

The results in that same table show Grok 4.7 at 46.3% on CursorBench 4.0 (against 40.4% for Grok 4.6, 41.7% for GPT-5.6 Sol and 51.8% for Fable 5.1); 71.0% on DeepSWE v1.1 in high-effort mode (against 65.2%, 72.7% and 70.0% respectively); 64.0% on EEBench, the electrical engineering benchmark (against 53.0%, 39.4% and 56.4%); 1,657 points on AA Briefcase v1.1, focused on multi-hour office work (against 1,546, 1,487 and 1,678); 37.6% on Terminal-Bench 4.0, covering extended terminal work (against 20.3%, 37.3% and 57.9%); 19.6% on the Harvey Legal Agent Benchmark, for legal work (against 15.8%, 2.5% and 6.7%); and 56.7% on HealthBench Professional, on clinical reasoning (against 48.5%, 60.5% and 62.1%). The announcement itself notes that Grok 4.7 does not lead across the board: Fable 5.1 beats it on several of these benchmarks, including CursorBench, DeepSWE, Terminal-Bench and HealthBench.

On professional knowledge tasks, xAI says Grok 4.7 improves its ability to create documents and presentations, and that on the GDPval and AA Briefcase benchmarks —which simulate work done by professionals such as lawyers, nurses and financial analysts— the model improves on Grok 4.6 and lands at a level comparable to other frontier models. In the GDPval Elo score comparison, Fable 5.1 (max mode) scores 1,735 points, Grok 4.7 (xhigh) 1,695, Grok 4.6 (high) 1,605 and GPT-6 Astra (max) 1,542.

On safety, the company stresses that Grok 4.7 was built with an entirely new set of safeguards and that it is, according to its own testing, the model most resistant to undue refusals and to jailbreak attempts. In dual-use domains such as cybersecurity and biological work, xAI maintains that the model combines usefulness for legitimate tasks with an appropriate refusal rate for dangerous uses, ranking first on LatchBio's biosecurity benchmark with 62.4%. On HackerBench v0.3, its internal test on risky cyber tasks, Grok 4.7 reportedly let through only 3.3% of risky dual-use requests while keeping a low block rate on legitimate security work. The company also mentions that it has begun granting limited, invitation-only access to certain cybersecurity partners so they can use the model's "red-team" capabilities for defensive research.

As for availability, Grok 4.7 can already be used in Cursor and Grok Build, as well as through the Grok API, third-party coding environments, and model routers and cloud platforms. The announced starting price is $2 per million input tokens and $6 per million output tokens, and xAI also offers a fast variant with double the output speed at double the price.

It is worth noting that the original material for this announcement comes from xAI itself, presented under the "SpaceXAI" brand on the source page, so the performance and safety figures cited are data published by the company about its own product and have not been independently verified in this summary.