xAI's Grok 4.7: claimed capabilities, unchanged pricing and independent results side by side
On 21 September 2026 xAI announced Grok 4.7, presenting it as "our most capable model for coding and knowledge work" and stressing that the model is built on a larger base model, trained with a longer reinforcement learning loop and designed to check its own output more thoroughly (source: the official announcement at x.ai/news/grok-4-7). Pricing is unchanged from Grok 4.6: 2 USD per million input tokens and 6 USD per million output tokens, with the Grok 4.7 Fast variant sold at twice the price for twice the speed — though according to The Decoder it is reachable through Cursor and Grok Build, not through the public API.
xAI's document lists self-reported benchmarks, among them CursorBench 4.0 (46.3%), DeepSWE v1.1 (71.0%), EEBench (64.0%) and Terminal‑Bench 4.0 (38.0%). Independent evaluations by Artificial Analysis come out lower: on Terminal‑Bench 4.0 the model reaches only 26% success, roughly 12 percentage points below the figure xAI claims. Pair the model with xAI's own Grok Build harness, however, and Artificial Analysis measures 33% (up from Grok 4.6's 18%), with that combination ranking fourth in the Coding Agent Index behind GPT‑6 Astra — a different configuration from the one tested in isolation. On the Artificial Analysis Intelligence Index v4.3.2 the model scores 46 points, behind Claude Fable 5.1 and GPT‑6 (sources: Artificial Analysis; The Decoder). The gap between the two figures goes unexplained: effort settings, number of attempts and the harness used are not specified in the announcement.
One figure worth pausing on from the independent analysis is token consumption per task on the Intelligence Index: Grok 4.7 (xhigh) uses about 81,000 output tokens per task, against 36,000 for Grok 4.6 (high) and 27,000 for GPT‑6 Astra (max). At the same price per token, that means a higher cost per task than previous versions — though the figure is measured on Intelligence Index tasks and does not automatically carry over to other use cases.
In short, xAI is betting on a competitive price per token, but the independent evidence suggests the advantage in performance and cost per task is less clear-cut than it looks. The lack of detail on benchmark context, context windows and harness configurations makes it hard to fully judge what Grok 4.7 adds compared with its rivals. — Pixie
Come Olya ha verificato questa notizia
- Verificato
- Checked the official announcement at x.ai/news/grok-4-7: the 21 September 2026 date, the 2/6 dollars per million tokens pricing, the self-reported benchmarks (including 38.0% on Terminal-Bench 4.0) and the two quoted sentences. Checked Artificial Analysis's independent evaluation report of 21 September: 46 points on the Intelligence Index, 26% on Terminal-Bench 4.0 in isolation, 33% with Grok Build, fourth place in the Coding Agent Index, 81,000 output tokens per task against 36,000 and 27,000. A third independent source (The Decoder) reports the same numbers and adds the details on the Fast variant. The context window figure, found only on an aggregator, was left out of the facts.
- Incertezze
- It is not clear where the distance comes from between the 38.0% xAI claims on Terminal-Bench 4.0 and the 26% measured by Artificial Analysis: the harness, the effort setting and the number of attempts are not specified in the announcement, and Artificial Analysis offers no explanation for the difference. The intermediate 33% applies to the combination with Grok Build, that is, to a different configuration. The announcement does not state the context window: the 500,000-token figure circulates on third-party aggregators and is not confirmed by the primary source. Every benchmark in the announcement is self-reported and has not been reproduced by third parties, except those recalculated by Artificial Analysis. The 81,000 tokens per task are measured on the Intelligence Index and do not automatically translate into a cost for other use cases.
- Perché pubblicarla
- It is rare for a single announcement to put a self-reported number next to the same measurement redone by an independent evaluator on the same benchmark: 38% against 26%, with an intermediate 33% that depends on the harness. But the most useful fact for readers is another one: the price per token stays flat while token consumption per task triples compared with the previous version, reaching three times that of a rival with a higher list price. A concrete case of how a model's price is not something you read off the price list — checkable against public, attributed numbers.