Grok 4.6 is official and rubs shoulders with the best. The real surprise is not the benchmarks, it is the price

Elon Musk had been promising it for three weeks, first "in a few weeks", then "probably next week". Well, today, August 12, it actually arrived: Grok 4.6 is available now, and SpaceXAI, xAI's new name following its integration into SpaceX, introduces it with a sentence that doubles as a mission statement: "frontier intelligence, a significant improvement over Grok 4.5 at the same price".

It ties Sol and sits one point behind Fable 5 Max

The numbers the company is showing are serious. On the Artificial Analysis index, the independent reference that condenses dozens of tests into a single figure, Grok 4.6 scores 61, level with GPT-5.6 Sol at maximum reasoning and just one point behind Claude Fable 5 Max, the smartest model on the market right now. A year ago, Grok was watching this race from the stands.

The gains come from a longer supplemental training run, stronger engineering data and expanded reinforcement learning, and they show up exactly where SpaceXAI wanted: agents that survive long tasks without losing the thread, and coding work in unfamiliar projects. The company reports improvements across every benchmark in that territory, from CursorBench to Terminal-Bench, and the model ships from day one on Cursor, OpenRouter, Vercel and Cloudflare, plus its own API.

The real headline is on the price tag

Here's the fist on the table: $2 to read a million tokens and $6 to write them, exactly what Grok 4.5 cost. For context, GPT-5.6 Sol charges $5 and $30 for the same job, and Claude Fable 5, $10 and $50. In other words, SpaceXAI is selling same-shelf intelligence at two to eight times less than its direct rivals.

There's a faster version that costs double. Even that one undercuts everyone.

The usual fine print still applies. That 61 measures what exams measure, not everyday experience, and Grok drags a history of outbursts its rivals don't have. But the message to the industry is hard to ignore: the frontier is no longer expensive by definition.

The price war just shifted up a gear

My bet is this moves pieces quickly. OpenAI and Anthropic have defended their premium rates by arguing nobody else reached their level; with a technical tie at a third of the price, that argument ages badly. And Musk has already promised Grok 4.7 within weeks, so the pressure won't let up.

One test remains that no benchmark can pass for you: using it. Once we've put real hours into the model, we'll tell you whether that 61 shows up in daily work or stays on the exam sheet. The announcement is today's; the verdict, a few days away.