What did Grok 4.7 actually fix?

Grok 4.7 is out, and the easy headline is "bigger model". True: it reportedly jumps to around 2.1 trillion parameters, up from about 1.5 trillion in 4.6. But that is not the interesting part. The interesting part is what xAI says it was trying to fix.

The stamina problem

Today's models are great at a single question. Ask one thing, get a sharp answer. But real work is not one question. Fixing a messy codebase, researching something properly, building a document across many steps: that can take hours. And over hours, models drift, lose track, forget what they were doing.

So xAI trained 4.7 with a longer reinforcement learning run on a harder mix of tasks, deliberately weighted toward problems that take many hours to finish. It is also meant to check its own work more carefully along the way. In other words: less sprinter, more marathon runner.

Did it work?

Some evidence says yes, and not just from xAI's own slides. On Artificial Analysis's independent coding agent index, the score went from 47 to 56. All three parts of that index improved, with the terminal tasks almost doubling. That puts Grok 4.7 fourth among the models tested in their own native setups, behind the top models from Anthropic and OpenAI. Not the winner. But a real step, not a cosmetic one.

The rest of the package: a 500,000 token context window, rebuilt safety filters that are meant to be harder to jailbreak while refusing fewer harmless requests, and the same price per token as before.

And why is Elon Musk's model suddenly sitting on Google's cloud?

Less dramatic than it sounds. Grok 4.7 is now available on Google's enterprise AI platform, but that platform is basically a giant store with over two hundred models in it, including models from Anthropic and Mistral. Grok just joined the shelf. Google makes money either way: customers pay to use the model, and it runs on Google's own infrastructure. Whoever wins the model race, the house still takes its cut.