Reader question: What changed in Grok 4.7, and what should a technically informed buyer learn from xAI’s price, speed, availability, and safety claims?
xAI’s September 21 release of Grok 4.7 follows a familiar frontier-model script: a larger base model, a longer reinforcement-learning run, and a table of benchmark scores. The more useful signal is the product framing around that table. xAI is selling Grok 4.7 as a model that can stay with difficult work longer, check itself more carefully, and do so at the same listed price and speed as Grok 4.6.
That is an efficiency argument. xAI says Grok 4.7 was trained on a harder mix of tasks weighted toward work that takes hours rather than seconds. It also says the model handles longer context and understands the Grok Bot harness natively. Those details point to a model aimed at sustained coding and knowledge work rather than a short-answer upgrade.
The price makes the positioning concrete. xAI lists starting rates of $2 per million input tokens and $6 per million output tokens. It says a fast variant doubles output speed at twice the price. The company’s headline says “twice as fast, at half the price of comparable models,” but that comparison is xAI’s framing; the page does not define one universal set of comparable models or workloads. The practical question is therefore not whether a token price looks low in isolation. It is whether a model completes a task with fewer retries, fewer tool steps, or less supervision.
The benchmark table should be read in that light. xAI reports 46.3% on CursorBench 4.0 and 71.0% on DeepSWE v1.1 at high effort, alongside results for office work, legal work, clinical reasoning, and electrical engineering. Those are useful clues about the tasks xAI prioritizes. They are not a substitute for an independent evaluation: the scores come from the vendor’s release page, use named effort settings, and compare with other vendors under the conditions xAI lists. A community reader can treat them as hypotheses about fit, not as a guarantee that every codebase or workflow will see the same result.
Safety is part of the pitch too. xAI says Grok 4.7 uses an entirely new safeguard stack and is its strongest model yet on refusal and jailbreak-resistance tests. It describes a balance between cyber-defence capability and refusal of risky dual-use requests, and says selected cybersecurity partners receive invite-only access to red-team capabilities. Those statements matter because a model aimed at long-running work can create more value when it is allowed to act, but also more risk when it acts on a mistaken instruction. The announcement does not establish how those safeguards perform in every deployment or jurisdiction.
Availability is broad on paper. xAI says Grok 4.7 is available in Cursor and Grok Build, and through the Grok API, third-party coding harnesses, model routers, and cloud platforms. That list tells readers the model is intended to travel across products, but it does not provide a complete matrix of regions, quotas, service terms, or third-party timelines.
The sensible decision rule is simple: compare completed work, not just advertised speed. Ask how often the model needs correction, how much context a task consumes, and what level of review a consequential output requires. Grok 4.7 may be a strong candidate when long tasks and cost discipline matter together. xAI’s announcement gives the reasons to investigate that fit; it does not prove the result for a particular team.
Source: xAI, “Introducing Grok 4.7”, September 21, 2026.