xAI launches Grok 4.7, built for coding and long-running tasks
Available from September 21 in Cursor and the API, the model keeps Grok 4.6's price while targeting complex projects.
Text assisted by artificial intelligence — reviewed by the author.
Translation of the original French article. Proposed by AI, reviewed by the author.
Abstract
SpaceXAI published the "Introducing Grok 4.7" announcement on September 21, 2026. The model, available the same day in Cursor, Grok Build and via the Grok API (identifier grok-4.7), remains at the same base price as Grok 4.6 (2 dollars per million input tokens, 6 dollars for output) and targets coding and knowledge work on long-running tasks. A Fast variant doubles output throughput at double the price. The API release notes confirm a context window of 500k tokens.
xAI put the Introducing Grok 4.7 post online at x.ai/news on Monday, September 21, 2026, accompanied by an update to the API release notes for the identifier grok-4.7. The announcement was published in the late afternoon, around 6:31 PM (Europe/Brussels) — as recently as the previous evening, nothing suggested this launch was coming.
According to the post, Grok 4.7 is presented as the company's most capable model for coding and knowledge work. It is said to hold up longer on difficult tasks, verify its own work more thoroughly, and come with a new stack of safeguards. It is offered at the same base price and speed as Grok 4.6, with a Fast variant offering twice the output throughput for twice the price.
The text describes a broader base than Grok 4.6, longer reinforcement learning training on a mix of harder tasks, weighted toward problems that take several hours, better handling of long context, as well as native training on the Grok Bot harness.
Availability and pricing, starting today
The model is available the same day in Cursor and Grok Build, via the Grok API, third-party coding harnesses, model routers, and cloud platforms.
The starting price cited in the post is 2 dollars per million input tokens and 6 dollars per million output tokens. The API release notes specify the product details: model grok-4.7, 500k context, text and image input, text output, with four reasoning effort levels — low, medium, high (default) and xhigh. On the API pricing side: 2 / 0.50 / 6 dollars per million (input / cache / output) under 200k prompt tokens, then 4 / 1 / 12 beyond that. Grok 4.7 Fast is reserved for Cursor and Grok Build; it is not available on the public xAI API.
What the in-house benchmarks show
The official table compares Grok 4.7 (xHigh), Grok 4.6 (High), GPT-5.6 Sol Max and Fable 5.1 Max. On CursorBench 4.0, Grok 4.7 reaches 46.3%, ahead of Grok 4.6 (40.4%) and GPT-5.6 Sol Max (41.7%), but behind Fable 5.1 Max (51.8%). On DeepSWE v1.1 at high effort, it posts 71.0%. On Terminal-Bench 4.0, the jump over Grok 4.6 is clear: 38.0% versus 20.3%. On EEBench, the score is 64.0%. On multi-hour office tasks (AA Briefcase v1.1) and clinical reasoning (HealthBench Professional), the post presents Grok 4.7 as a clear improvement over 4.6, within the pack of frontier models.
These scores remain those from the announcement and will need to be cross-checked with independent evaluations once the model is widely accessible outside xAI channels.
Safeguards, cyber and biosecurity
xAI claims a new safeguards stack, better refusals and increased resistance to jailbreaks. On its HackerBench v0.3 benchmark, the post claims that only 3.3% of risky dual-use prompts get through, while limiting false blocks on legitimate security work. An invite-only access to red-team capabilities is open to selected cybersecurity partners. On the LatchBio biosecurity benchmark, the text claims 62.4%.
These claims remain vendor claims: they do not replace a public third-party audit of production behavior.
The next step will be operational: API latency and stability outside of demos, real-world performance on CursorBench or SWE beyond in-house scores, and how cyber safeguards behave as traffic scales up.
Sources
- Introducing Grok 4.7 — SpaceXAI / x.ai, September 21, 2026
- Release notes — docs.x.ai (Grok 4.7 /
grok-4.7entry), September 2026 - News index — x.ai/news, entry dated September 21, 2026
Frequently asked questions
What is Grok 4.7
It is the new frontier model announced by SpaceXAI on September 21, 2026, geared toward coding and knowledge work on long-running tasks, with a safeguards stack presented as renewed.
Is Grok 4.7 available today
Yes, according to the announcement, via Cursor, Grok Build, the Grok API (grok-4.7), third-party harnesses, and cloud platforms. The Fast variant remains limited to Cursor and Grok Build.
What is the price of Grok 4.7
The post cites $2 per MTok for input and $6 per MTok for output for the base version, at the same level as Grok 4.6. The API release notes also detail the cache price and the doubling of prices beyond 200k prompt tokens. The Fast version doubles both throughput and price.
How does Grok 4.7 differ from Grok 4.6
According to xAI: a broader base, longer RL on multi-hour tasks, better self-verification and improved context handling, native training on Grok Bot, and new safeguards. The official tables show gains on several benchmarks, notably on Terminal-Bench and CursorBench versus 4.6.
The AI Desk. (2026). xAI launches Grok 4.7, built for coding and long-running tasks. The AI Desk. https://ntilia.com/u/aidesk/en/xai-launches-grok-4-7-built-for-coding-and-long-running-tasks (consulté le 2026-10-22)