xAI's Grok 4.7 arrived on OpenRouter on 21 September 2026, five weeks after Grok 4.6. It is an incremental release, and every number moves the right way.
Side by side
Prices are list prices on xAI's own endpoint. Speed figures are OpenRouter's median over a trailing thirty-minute window on that endpoint, read 23 September 2026. Index scores come from each model's Artificial Analysis page.
The three changes
Price: 20% off across the board. Input, output and cached input all drop by a fifth. At 10M output tokens a month, the bill goes from $60 to $48.
Speed: faster at both ends. Throughput rises from 48 to 61 tok/s, and the first token arrives in 1.29 s instead of 1.75 s.
Quality: two points. A move from 44 to 46 is small on a composite index, and it is within the range where task-specific results can disagree. It does put Grok 4.7 level with Xiaomi's MiMo-V2.6-Pro, released the same day, and two points below GPT-6 Sol.
What did not change
The unusual output ceiling stays. Grok 4.7 accepts 450K output tokens in a single request, against 128K for Claude Opus 5.5 and GPT-6 Sol. If your workload generates long documents or large code files in one pass, that is still Grok's clearest structural advantage. The endpoint takes text, images and files, and exposes a reasoning_effort parameter.
Should you switch?
If you are on Grok 4.6, yes. The model is cheaper, faster and no worse on anything we can source, and it keeps the same context and output limits. If you are choosing fresh, the question is whether 450K output tokens matter to you. If they do not, MiMo-V2.6-Pro matches Grok 4.7's index score at a fifth of the output price, though it is less than half as fast. Compare the two directly.
Our category ratings for both Grok models are editorial estimates anchored to the Intelligence Index, labelled Estimated, until accepted benchmark results replace them.
