xAI launches Grok 4.7 with upgraded coding at half the price

If you use AI to code every day, you already know the sinking feeling when a model completely forgets what it was doing three hours into a complex refactor.

xAI is taking direct aim at that fatigue. The company just released Grok 4.7, billing it as their top-tier engine for serious coding and multi-hour knowledge work. It runs at the exact same price as Grok 4.6, coming in at $2 per million input tokens and $6 per million output tokens, alongside a fast version that doubles both the speed and the price tag.

Built for marathon tasks

Instead of just tweaking prompts, xAI swapped in a larger base model and stretched out reinforcement learning across harder, multi-hour problems.

The pitch is that Grok 4.7 is much better at catching its own mistakes before finishing a response. On coding-specific tests like CursorBench 4.0, which targets long-running workflows, the model put up a 46.3% score. It also logged 64.0% on EEBench and showed noticeable gains on professional suites like GDPval and AA Briefcase, where it mimics work done by lawyers and financial analysts.

Safety got an overhaul as well. xAI introduced a fresh safeguard stack aimed at dangerous dual-use scenarios. It topped LatchBio's biosafety benchmark at 62.4% and allowed only 3.3% of risky malicious prompts through on HackerBench v0.3, all while trying not to trip up legitimate security researchers.

It is available right now inside Cursor, Grok Build, third-party harnesses, and via the xAI API.

The benchmark debate

While xAI claims frontier price-performance, independent reviewers are already raising eyebrows.

Tech reviewers looking at early numbers from Artificial Analysis noted that Grok 4.7's overall intelligence index still sits behind frontier models like GPT-6 and the latest Claude lineup. In fact, on some composite indexes, it merely ties open-weight models like Mimo 2.6 Pro.

More interestingly, some evaluators pointed out that while token rates look cheap on paper, the effective cost per completed task on complex setups can actually climb higher than running GPT-6, muddying xAI's half-price narrative.

Why it matters

We see Grok 4.7 as a clear bid to capture developer mindshare where it actually happens: inside tools like Cursor.

For engineers, this model is worth testing immediately if you work on large codebases that routinely derail standard LLMs. A model trained specifically on long-horizon self-verification could shave hours off mundane debugging, and the barrier to try it is effectively zero if you already use Cursor.

However, xAI has a narrative hurdle to clear. If independent benchmarks continue to show that actual task completion costs are higher than advertised, developer enthusiasm will cool off fast. Low token prices mean nothing if a model takes twice as many attempts to get a pull request right.

Still, by pushing Grok 4.7 straight into production coding harnesses on day one, xAI is ensuring competitors cannot treat coding benchmarks as an afterthought anymore.