SpaceXAI: Grok 4.6

This release moved the scoreboard without moving anything a buyer pays for. In the catalog, as observed on 31 August 2026, this row and x-ai/grok-4.5 listed the same context window — 500000tokensopenrouter-models, last checked 2026-09-14 there against 500000tokensopenrouter-models, last checked 2026-09-14 here — the same input price, $2.00per million tokensopenrouter-models, last checked 2026-09-14 against $2.00per million tokensopenrouter-models, last checked 2026-09-14, and the same output price, $6.00per million tokensopenrouter-models, last checked 2026-09-14 against $6.00per million tokensopenrouter-models, last checked 2026-09-14.

What separates them is measurement. Artificial Analysis's intelligence index reads 39.1openrouter-models, last checked 2026-09-14 on the July row and 44.4openrouter-models, last checked 2026-09-14 on this one, five weeks later, with this row's coding index at 76.8openrouter-models, last checked 2026-09-14 and its agentic index at 53.4openrouter-models, last checked 2026-09-14. Same window, same meter, higher numbers.

The envelope moved a release earlier, and so did the reasoning settings this row is often credited with. x-ai/grok-4.3 listed 1000000tokensopenrouter-models, last checked 2026-09-14 of context at $1.25per million tokensopenrouter-models, last checked 2026-09-14 input, and read falseopenrouter-models, last checked 2026-09-14 for reasoning.mandatory at an effort of lowopenrouter-models, last checked 2026-09-14; this row reads trueopenrouter-models, last checked 2026-09-14 at highopenrouter-models, last checked 2026-09-14. Grok 4.5 is where that flipped — in the same snapshot it already carries this row's pair, along with the halved window and the raised price — and the only reasoning difference left between the two is an extra xhigh effort tier here. Anyone dating the end of optional reasoning to 2026-08-12source, accessed 2026-08-28 is a month late.