xAI releases Grok 4.7 as its new flagship text model

A flagship version bump

xAI officially released Grok 4.7 on September 21, 2026, as its primary flagship text model. Unlike a brand-new product name that might signal a complete shift in assumptions, the 4.7 identifier indicates a version bump on an existing ladder. This means that most of the integration guidance, tooling, and evaluation harnesses used for previous versions will carry over to this release.

The model arrived without any companion models, such as a vision variant or a smaller sibling model. This distinction is important because many AI launches arrive in families where one model is retired or replaced by a specialized variant. With Grok 4.7, the release is focused entirely on this single flagship model rather than a suite of new tools.

Pricing for the model is set at $2 for input and $6 for output per million tokens. These figures are provided at the list rate with no markup, which means testing the model against its own release date costs only the tokens it uses. This transparency allows for straightforward cost calculations for users looking to integrate the model into their workflows.

The crowded release window

The timing of the Grok 4.7 launch is a significant factor in how the model is being perceived by the public. It arrived one day before a day that saw three other flagship model releases from two major labs. This creates a situation where measurement capacity and public attention are split across several high-profile launches simultaneously.

Because independent evaluators are not infinite, a model released in a crowded window often receives a thinner public record. This is not necessarily a reflection of the model's quality, but rather a result of the schedule. When multiple labs release flagship models in the same week, the attention that pays for measurement goes to the date rather than the specific model.

As a result, current records for Grok 4.7 are somewhat uneven. On October 7, 2026, the model only had 8 rows of data across 3 categories in some independent harnesses. This means the available data is a map of what has been measured so far, rather than a comprehensive summary of the model's total capabilities.

Technical configurations

When comparing Grok 4.7 to other models like those from OpenAI or Anthropic, users must be careful to check the specific settings used. For example, Grok 4.7 was measured at the Xhigh reasoning setting on certain independent harnesses. In contrast, some competitors in the same comparison tables were measured at Max settings.

Comparing these numbers side-by-side without noting the different configurations can lead to misleading conclusions. A model that is not ranked in a specific table is not necessarily a model that ranked poorly; it simply means it was ineligible for that specific ranking at the time of the reading. Users should look for the specific category results that match their own workload requirements.

The data currently shows that the model has an Intelligence Index of 46.45 based on certain live pages. However, because these figures are taken from a reading date of October 7, 2026, they are considered moving targets. Benchmarks have a shorter shelf life than the interval on which most websites are refreshed, meaning scores can change as harnesses are re-run.

Navigating the data

To accurately assess Grok 4.7, users should prioritize the identifier over the prose in any news report. Since it is a version bump, any page describing it as a completely new line might have a positioning problem. Identifying it as a successor rung helps users understand how much of their existing integration will carry over to the new model.

It is also important to distinguish between facts about the model and facts about the calendar. For instance, saying a model shipped during a crowded week is a fact about the schedule, not a statement on the model's performance. Keeping these two facts separate prevents the compression of information that can lead to a misunderstanding of the model's actual capabilities.

Finally, users should attach a specific read date to every score they find online. Because a model can have different figures recorded on different days, the spread between those numbers can be larger than the difference between two different models. Treating the ship date as the only fixed fact while treating reading dates as moving ones is the most reliable way to track the model's progress.