The new model reaches parity with GPT-5.6 Sol on the Artificial Analysis benchmark of benchmarks, going from ninth to fourth on the leaderboard, and it also beats Sol on several benches on the Grok presentation.
SpaceXAI claims this proves «frontier level performance,» and it does come in striking distance to the latest state-of-the-art model Opus 5. Considering that GPT Astra is due shortly, this might be a short-lived claim, however.
The new Grok was trained on a «wide range of agentic tasks,» not to mention the corpus of real-word sessions from Cursor, with a focus on knowledge work and general coding.
It is supposedly especially good at «turning a broad product idea into a working first version,» which can then be iterated upon. It is also proficient on long-running, complex tasks and creative visual work, SpaceXAI says.
The main selling point of the model is undoubtedly going to be price/performance, starting at $2 per million input tokens and $6 out, it is astoundingly cheap to run. Prices for GPT-5.6 Sol are $5 for input and $30 for output, which works out at more than twice the input price and 5x on output. Grok 4.6 also has a faster version at double the cost.
The model is out now, and can be found on Cursor, Grok Build and in the API.
Read more: SpaceXAI’s presentation and launch post, Musk on costs, Artificial Analysis rundown, Gizmodo. Discussion on Hacker News and r/Singularity.