Where it stands
xAI has officially announced Grok 4.7 as its model for longer coding and knowledge-work tasks, listing $2 per million input tokens and $6 per million output tokens. Its benchmark and safety results are company-reported and have not been independently tested. Before the announcement, Elon Musk's 10-day countdown pointed to about 12 September, and the model was spotted in OpenCode Zen and the Grok app. Rate limits, billing terms and regional availability remain open, and reports about a successor, Grok 4.8, are still preparatory.
Grok 4.7 appears in the Grok app's Expert mode
An X post from @LuminaBench reports that Grok 4.7 is now in the Grok app. The attached screenshot shows the app’s mode selector with Expert selected and a tooltip that reads “Thinks hard Grok 4.7.”
The post indicates that Grok 4.7 has appeared in at least one app experience, but it does not establish a broad or universal rollout. It also does not clarify whether access is part of a limited test, a staged release or general availability.
What the report says about Grok 4.7 in the app
The post gives a short update: “Grok 4.7 is now in app.” Its accompanying reply says that models should be released in the app on the same day they launch. That comment expresses frustration about release timing, but it does not provide a release date or additional details about access.
What the screenshot shows
The screenshot shows a dark app interface with a mode menu listing Auto, Fast, Expert, Build and Heavy. Expert is highlighted as the selected option. A nearby tooltip reads “Thinks hard Grok 4.7.”
This supports the narrower interpretation that the app interface displayed a Grok 4.7 label alongside the Expert mode. It does not show what the modes do, whether Expert is exclusive to Grok 4.7 or whether selecting it changes performance in a particular way.
Image credit: @LuminaBench on X
What the evidence does—and does not—confirm
The post and screenshot support a report that Grok 4.7 has appeared in an app interface. They do not confirm:
that every Grok user can access it;
whether the rollout is limited, staged or broad;
when the model became available;
whether access depends on a subscription or particular app version;
what Grok 4.7 can do;
how it performs; or
whether it is available outside the app.
The screenshot also cannot establish whether the Expert, Build or Heavy modes represent different models, different processing settings or another app-level distinction.
For readers tracking Grok 4.7 app availability, the practical takeaway is limited: the model name is visible in the reported app screenshot, but the post does not establish universal access or explain the model’s capabilities.
Sources
Grok 4.8 identifiers reportedly appear in Cursor's model list
Grok 4.8 has reportedly appeared in Cursor’s server-side model list, alongside identifiers that look like effort or speed variants. In a post on X, @LuminaBench attributed the sighting to a Magpie commit and described it as preparation for the next Grok release. Source on X: https://x.com/LuminaBench/status/2104505848223461846
The report does not show that Grok 4.8 is available in Cursor. It also does not establish a release date or prove that users can select any of the listed identifiers.
What was reportedly spotted in Cursor’s model list
LuminaBench said the server-side list includes grok-4.8, along with three related CLI-style identifiers:
grok-4.8-highgrok-4.8-high-fastgrok-4.8-xhigh-fast
The screenshot describes a distinction between the broader grok-4.8 family name and a CLI request for an effort or speed variant. That supports reading the names as model-related labels, but it does not establish how the variants work or whether they are exposed to Cursor users.
Image credit: @LuminaBench on X
What the reported variants show
The names appear to indicate different effort or speed configurations, although the source does not explain their exact behavior. They should therefore be treated as reported identifiers rather than confirmed user-facing modes.
The listing also provides no supported information about the variants’ capabilities, performance, pricing or access rules. Their appearance does not by itself show that the model is finished, selectable or operating in a public Cursor integration.
Is Grok 4.8 available yet?
No availability conclusion follows from this sighting. LuminaBench repeatedly clarified that the entry was a sighting of preparation, not evidence that Grok 4.8 had launched or that its release was close.
A server-side model list can be consistent with a company preparing support for a model or reserving names for future use. On its own, however, it cannot establish public access, a completed integration or a release announcement from xAI.
Does the listing indicate when Grok 4.8 will launch?
It does not provide a reliable timeline. LuminaBench gave personal guesses of weeks to about a month, while repeatedly saying the listing looked like preparation rather than a signal of an imminent launch.
Those comments are speculation, not a release schedule. Based on the reported listing alone, the most that can be said is that Grok 4.8-related identifiers were reportedly present in Cursor’s server-side preparation. The timing and public availability remain unconfirmed.
Sources
xAI announces Grok 4.7 for long coding and knowledge-work tasks
xAI describes Grok 4.7 as its most capable model for coding and knowledge work, with a focus on tasks that run for hours rather than short, isolated prompts. The announcement says the model uses a larger base model than Grok 4.6, underwent a longer reinforcement-learning training run, and is better at checking its own work and managing longer context. xAI also lists several access routes.
The results and safety figures below are reported by xAI. The announcement does not provide independent testing, detailed methodologies, sample sizes, confidence intervals, or evidence that the results will transfer to every user’s workload.
What Grok 4.7 is designed to do
Grok 4.7 is intended for longer-running coding, terminal, document, presentation, legal, clinical, and engineering tasks. xAI says it trained the model on a harder mix of problems weighted toward tasks that take many hours to complete. It also says Grok 4.7 was trained to natively understand the Grok Bot harness, which is meant to improve conversational tasks and broader knowledge work.
The practical emphasis is therefore less about a single new feature and more about persistence: continuing through a complex task, reviewing intermediate work, and handling a larger amount of context. Those are design goals described by xAI, not a guarantee that the model will complete every long-running task successfully.
How Grok 4.7 compares with Grok 4.6 and other models
xAI’s announcement compares Grok 4.7 with Grok 4.6 and two other named models across coding, office work, terminal tasks, legal work, clinical reasoning, and electrical engineering. The published figures are reported results rather than a universal ranking. The source does not explain every benchmark’s scoring scale or provide the full test conditions.
Benchmark or task area | Grok 4.7 (xhigh) | Grok 4.6 (high) | GPT-5.6 Sol (max) | Fable 5.1 (max) |
|---|---|---|---|---|
CursorBench 4.0, software engineering | 46.3% | 40.4% | 41.7% | 51.8% |
DeepSWE v1.1, software engineering | 71.0%* | 65.2% | 72.7% | 70.0% |
AA Briefcase v1.1, multi-hour office work | 1,657 | 1,546 | 1,487 | 1,678 |
Terminal-Bench 4.0, multi-hour terminal work | 38.0% | 20.3% | 37.3% | 57.9% |
Harvey Legal Agent Benchmark | 19.6% | 15.8% | 2.5% | 6.7% |
HealthBench Professional | 56.7% | 48.5% | 60.5% | 62.1% |
EEBench, electrical engineering | 64.0% | 53.0% | 39.4% | 56.4% |
*The asterisk marks Grok 4.7’s high-effort DeepSWE result, according to xAI.
The xhigh and high labels indicate the listed effort settings for Grok 4.7 and Grok 4.6. Those settings matter when interpreting both performance and price comparisons: a result may involve a different amount of model effort, and the table should not automatically be read as a comparison of identical configurations.
On the reported figures, Grok 4.7 improves on Grok 4.6 in every listed comparison. Among the percentage-based comparisons, its largest increase is on Terminal-Bench 4.0, rising from 20.3% to 38.0%. That percentage-point change should not be compared directly with the raw-score change on AA Briefcase, which uses a different scale.
The figures also do not make Grok 4.7 the top model in every category shown. Fable 5.1 has the highest listed scores on CursorBench, AA Briefcase, Terminal-Bench, and HealthBench Professional. GPT-5.6 Sol is higher on DeepSWE and HealthBench Professional, while Grok 4.7 leads the listed Harvey Legal Agent Benchmark and EEBench comparisons.
Coding and terminal work
The coding results support xAI’s emphasis on longer tasks, but the benchmarks evaluate different kinds of work. CursorBench 4.0 is presented as a test of longer-running software-engineering tasks, while Terminal-Bench 4.0 focuses on multi-hour terminal work. A strong result on one does not necessarily predict the same result on the other.
DeepSWE v1.1 adds another qualification because Grok 4.7’s listed score is marked as high effort. Without the benchmark methodology and effort settings for every model, these results are best read as xAI’s reported comparison rather than as a complete independent performance test.
Documents and professional knowledge work
xAI says Grok 4.7 is better at creating documents and presentations. It cites GDPval and AA Briefcase as evaluations involving tasks performed by professionals such as lawyers, nurses, and financial analysts.
The announcement says Grok 4.7 improves on Grok 4.6 on both benchmarks and performs comparably with other frontier models. For a potential user, that supports an intended use case around multi-step professional work, not a verified claim that Grok 4.7 can replace a lawyer, nurse, financial analyst, or other professional. Human review remains important for consequential documents and decisions.
Safety, refusals, and cybersecurity
xAI says Grok 4.7 uses a new safeguard stack and is the strongest model it has tested for refusal behavior and jailbreak resistance. In biological and cybersecurity scenarios, the company presents the model as aiming to preserve usefulness for benign tasks while refusing dangerous requests.
The announcement reports a 62.4% result on LatchBio’s biosafety benchmark and says Grok 4.7 topped that evaluation. It also reports that the model allowed 3.3% of risky dual-use prompts through on HackerBench v0.3, which xAI describes as its benchmark for risky and malicious cyber tasks. The company says the model rarely blocks legitimate security work and has begun offering selected cybersecurity partners invite-only access to its red-team capabilities for defense research.
These are company-reported safety results, not a complete assessment of the model’s safeguards. The announcement does not describe the benchmark samples, scoring procedures, or the definition of legitimate security work in enough detail to independently interpret the figures. The 3.3% figure also should not be read as a general error rate for all cybersecurity prompts.
Grok 4.7 pricing and availability
xAI lists Grok 4.7 at a starting price of $2 per million input tokens and $6 per million output tokens. It also says a fast variant provides twice the output speed at twice the price. The announcement does not state the fast variant’s exact output speed or provide API rate limits, billing terms, context-window details, regional availability, or service-level commitments.
The source lists Grok 4.7 as available through:
Cursor
Grok Build
The Grok API
Third-party coding harnesses
Model routers
Cloud platforms
xAI also says users can try Grok Build for free. Access-channel details may differ, including the available variant, limits, pricing, and performance, so users should check the terms for the route they plan to use.
The announcement presents an apparent pricing inconsistency. Its opening description says Grok 4.7 is served at the same price and speed as Grok 4.6, while the comparison table lists Grok 4.7 at $2 per million input tokens and $6 per million output tokens, versus $4 and $20 for Grok 4.6. The table therefore shows lower listed token prices for Grok 4.7, but the different xhigh and high labels mean the figures should not automatically be treated as identical configurations.
For API buyers, input and output token prices are only part of the decision. Total cost will also depend on how much context a task uses, how many attempts or verification passes it requires, whether the fast variant is selected, and the limits imposed by the chosen access channel. The announcement does not provide enough information to calculate a typical task cost.
What the announcement does—and does not—show
Grok 4.7 is presented as a model for extended coding and knowledge-work tasks, with a reported improvement over Grok 4.6 in the comparisons shown. xAI also positions it as a model that can combine stronger performance on legitimate cybersecurity work with more selective refusals for dangerous requests.
The announcement does not include independent hands-on testing, benchmark methodology, real-world examples, or results from typical user workloads. Readers evaluating Grok 4.7 should therefore treat the reported scores as signals about xAI’s intended positioning, then check the relevant access channel’s pricing, limits, and performance on their own tasks before committing to it.
Sources
Grok 4.7 spotted in the OpenCode Zen harness
A post on X reports that Grok 4.7 has appeared in the OpenCode Zen harness, but there was no official release or public access at that point.
In a post on X, @Conor_D_Dart says that “Grok 4.7 is live in a harness” and that it was “spotted in the Opencode Zen harness.” The post also says the model is “arriving like now,” framing the sighting as a possible sign of an imminent arrival rather than providing a confirmed launch announcement.
What the post reported
The specific observation was that Grok 4.7 had been seen in OpenCode Zen, which the post describes as a harness. In general, a harness is a software environment used to connect to, test or run a model.
The excerpt also describes Grok 4.7 as live in a harness. That wording does not establish that the model is generally available to users. A model can appear in a testing, integration or limited-access environment without having a public release, published documentation or an announced sign-up process.
What the sighting does not establish
At that point it wasn't clear:
An official release statement
A public release date or confirmation that the model has launched
Instructions for accessing Grok 4.7
Information about its capabilities or performance
Pricing or availability details
Independent verification of the OpenCode Zen sighting
An explanation of how Grok 4.7 was identified
So the OpenCode Zen reference was a reported sighting, not confirmation that Grok 4.7 was available there or anywhere else. It also does not establish whether the appearance was a test, a limited integration or a broader release.
For now, the report supports monitoring Grok 4.7 for a formal announcement or additional evidence. It does not support conclusions about the model’s quality, release timing or access.
Sources
Musk says Grok 4.8 training is nearing its end
In a post on X, Elon Musk (@elonmusk) says Grok 4.8 is a 2.5T model trained with a new C++ software stack. He says its training will finish this week, after which the model will start RL. The post does not explain what “RL” refers to in this training plan.
What Musk’s post says about Grok 4.8
The post provides four details:
Grok 4.8 is described as a 2.5T model.
It was trained with a new C++ software stack.
Its training is expected to finish this week.
It is then expected to start RL.
The post does not say whether training has since finished.
What the 2.5T size and C++ stack mean here
The post gives a reported size for Grok 4.8 and identifies the software stack used during training. It does not say whether 2.5T refers to the number of parameters or another measure of model size. It also does not provide technical details about the C++ stack or describe how the stack affects training.
Those details do not establish how Grok 4.8 will perform. The post includes no benchmark results, capability descriptions or comparison with earlier models.
What happens after training, according to the post
Musk says Grok 4.8 will start RL after its current training phase. The post does not define what “RL” means in this particular plan, so it does not establish the methods, objectives or duration of that stage.
It also does not say when that stage will end or whether it will be followed by a public release.
What remains unknown about Grok 4.8
Still unknown at that point:
Grok 4.8’s capabilities or benchmark performance
A launch date or availability
How the new C++ software stack is designed or used
What “RL” means in this training plan
Whether the reported 2.5T size has been independently verified
For now, the update is a narrowly scoped description of Grok 4.8’s reported size, training software and expected next training phase—not a capability announcement or release confirmation.
Musk gives a 10-day countdown for Grok 4.7
Grok 4.7 may be released around September 12, 2026, based on a September 2 post from @elonmusk saying that the new model “comes out in 10 days.” The announcement gives a release countdown but does not provide further information about the model itself.
What the Grok 4.7 announcement says
The post is a brief timing announcement: Grok 4.7 is expected to come out 10 days after the post was published. Counting forward from September 2 produces an estimated date of September 12, 2026.
That date should be treated as an estimate based on the wording of the announcement rather than a separately stated calendar launch date. The post does not identify a specific release time or explain whether availability will begin simultaneously for all users.
What is not yet established
The announcement does not describe Grok 4.7’s capabilities, performance, pricing, access requirements, or supported products. As a result, the confirmed news is limited to the stated release timing. Any claims about new features or improvements would require additional information from a later announcement or release documentation.
For now, the key update is that Grok 4.7 has been given a near-term release window, with September 12 the approximate date implied by the post.
0 comments
No approved comments yet. You can start the conversation.
Leave a comment