Grok 4.7 has been released as a notable improvement over Grok 4.6, maintaining the same price and speed. It works longer on difficult tasks, checks its work more carefully, and includes the strongest safeguards to date. Grok 4.7 is available now in Cursor, Grok Build, and the Grok API.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
xAI released Grok 4.7 on September 21, 2026, officially stating it is a notable improvement over Grok 4.6 at the same price and speed, and emphasizing longer work on hard problems, more careful self-checking, and the strongest safety protections to date. But all performance claims come from the publisher's own statements, lacking third-party verification and specific benchmark data.
On September 21, 2026, xAI announced the release of Grok 4.7 via SpaceXAI's X account, with the core statement being just one line: it is a "notable improvement" over Grok 4.6, and "same price and speed." The post received about 19.7 million views, indicating no small amount of market attention, but the official announcement did not include any benchmark scores, evaluation methodology, or third-party comparison data.
The same account then added three capability descriptions: Grok 4.7 works longer on difficult tasks, checks its own work more carefully, and comes with "the strongest safety protections to date." These are directional descriptions rather than quantifiable metrics—no indication of how many steps or tokens "longer" means, nor what the safety protections are strong at or who evaluated them. Therefore, what can be confirmed from this release is the product's existence, launch time, and channels; conclusions at the performance level can currently only be considered publisher claims.
In the absence of benchmark scores, "same price and speed" is actually the most specific and verifiable piece of information in this release. It limits Grok 4.7's positioning to: the improvement does not come at the cost of a price increase or sacrificing response speed. For developers already using Grok 4.6, this lowers the decision cost of migration—if the official statement holds, upgrading does not require redoing cost and latency budgets.
But this constraint is also only from the official line. The source does not provide specific price figures, throughput or latency metrics, nor independent measurements. Therefore, "same price and speed" should be understood as the publisher's product promise, not verified operational parameters. What can truly verify it are subsequent API bills and actual call latency, and such data does not exist in this source.
Officially, Grok 4.7 is "now available," with launch channels including Cursor, Grok Build, and Grok API. The simultaneous opening of three channels means this release is oriented toward developers and coding scenarios, rather than only general chat users. Combined with the official demo choosing the task of "building an open-world city game," it can be inferred (this is an inference) that xAI wants to anchor the improvement narrative of 4.7 on long-horizon, multi-step code and engineering tasks.
It is necessary to distinguish: channel availability is a fact explicitly stated by the official, while "improvements mainly oriented toward coding scenarios" is a judgment derived from the demo topic and channel combination, and the source does not directly state this. The source also does not explain the availability scope of each channel, regional restrictions, or whether it is a phased rollout.
The official comparison material is Grok 4.7 (first) and Grok 4.6 (second) building the same open-world city game. Such demos are intuitive, but they do not constitute performance evidence by themselves: no scoring criteria, no number of repetitions, no explanation of whether the parameters of the two runs were consistent.
This weak point was immediately pointed out by users. A user publicly questioned: why compare Grok 4.7's xhigh setting with Grok 4.6's high setting, rather than using xhigh for both. The source does not contain any response from xAI to this question. This means the conclusion of "notable improvement" is currently built on a set of official demos with possibly unequal settings, and readers should treat it as a claim to be verified, not an established conclusion.
Separate the confirmed from the unconfirmed: what is confirmed is the release time (September 21, 2026), product name, launch channels, and the official statements about same price and speed, longer task handling, more careful self-checking, and strongest safety protections. What is unconfirmed is the actual magnitude of these improvements, the specific content of the safety protections, and whether they hold up in independent testing.
For developers, the pragmatic approach is: since the official claims price and speed are unchanged, you can directly compare 4.6 and 4.7 on your own real tasks, especially long-step tasks; at the same time, pay attention to the setting levels during comparison to avoid repeating the possible setting inequality in the official demo. Before third-party benchmarks or independent evaluations appear, any quantitative conclusion about "notable improvement" lacks source support.
All information in this article comes from a single source published by xAI via SpaceXAI's X account (source_tier 3, press_release). The product release, launch channels, same price and speed, and capability descriptions are all publisher self-statements, belonging to source_claim; the source did not provide benchmark scores, price figures, latency metrics, or third-party verification. The user's question about unequal comparison settings comes from comments included in that source, and xAI did not respond in the source. The judgment in the article that "improvements are oriented toward coding scenarios" is an inference and has been clearly marked.
The release of Grok 4.7 confirms the product, time, and channels, and makes "same price and speed" the hardest promise; but "notable improvement" is currently supported only by official demos and self-statements, and the demo has a public question about unequal settings. Before independent evaluations appear, it is better treated as an upgrade to be verified, rather than an established performance leap.
Primary report
Primary source