Grok 4.7 is not released. As of 12 September 2026, xAI’s developer release notes contain no grok-4.7 model ID, no pricing, no context window and no benchmark card.
That is the checkable part, and it is the part most coverage skips. Every date this model has ever had came from Elon Musk’s personal posts on X. xAI, the company, has published nothing: no launch post, no model page, no specification sheet.
This separates what is documented from what is claimed, because on Grok 4.7 those are two very different lists.
In this guide:
- What Is Actually Confirmed
- The Five Missed Windows
- What Musk Has Claimed
- Grok 4.6: The Model That Exists
- The SpaceX Data Angle
- Why It Slipped, In Musk’s Own Words
- How to Check for Yourself
- Frequently Asked Questions About Grok 4.7
- Verdict
- Sources and Further Reading
What Is Actually Confirmed
The complete documented record, as of 12 September 2026:
| Claim | Status |
|---|---|
| Grok 4.7 exists as a project | Confirmed by xAI leadership statements |
| Release date | None published by xAI |
| API model ID | Does not exist |
| Pricing | Not published |
| Context window | Not published |
| Benchmark card | Not published |
| Parameter count | Founder claim only |
| SpaceX training data | Founder claim only |
Four independent checks of xAI’s news page, model directory, API release notes and llms.txt on 11 and 12 September found no reference to Grok 4.7 in any of them. The newest model in xAI’s own documentation is Grok 4.6.
That is not a gap in reporting. It is the entire confirmed record.
The Five Missed Windows
Every date associated with Grok 4.7 traces to a Musk post on X. None appeared on x.ai.
| Date posted | What was said | Implied window |
|---|---|---|
| 24 July 2026 | “in 4 weeks” | Late August |
| 28 July 2026 | “a few weeks” after Grok 4.6 | Late August |
| 12 August 2026 | “3 to 4 weeks” | Early September |
| 1 September 2026 | “comes out in 10 days” | ~11 September |
| 11 September 2026 | “needs a few more days to cook” | Unspecified |
Five statements, five windows, none met. The 11 September post is the most recent and the only one that concedes the previous target passed.
Community trackers described the release as “officially postponed” around 11 and 12 September. Treat that as characterisation rather than fact: xAI has published nothing about a delay, just as it published nothing about a date. There is no official schedule to postpone.
The pattern matters more than any individual slip. A release that has moved five times on founder statements alone, with zero company documentation at any point, is not a release with a date. It is a project with commentary.
What Musk Has Claimed
For completeness, the specifications attributed to Grok 4.7, all from Musk posts on X between late July and mid-August 2026, none from xAI documentation:
2.1 trillion parameters, against Grok 4.6’s 1.5 trillion. That would be roughly a 40% increase.
“Better than 4.6 in every way, except slightly slower to serve.” An unusual trade to advertise, and it points at a larger model rather than a more efficient one.
“Even better token efficiency,” meaning fewer output tokens per completed task. If that holds, it matters more in practice than a benchmark win, because cost per task tracks tokens rather than rate cards. The same argument appears in our GPT-6 Astra coverage, where token efficiency turned out to be the real cost differentiator.
Supplemental training on SpaceX company data.
Every item on that list is a company-described specification rather than a documentation entry. None has a model card behind it.
Grok 4.6: The Model That Exists
Worth covering properly, because it is what you can actually use today and it is genuinely competitive.
Grok 4.6 shipped on 12 August 2026 on a 1.5-trillion-parameter foundation. It offers a 500K context window, accepts text and image inputs, and has no stated text output limit. Pricing runs around $2.00 per million input tokens and $6.00 output, with cached input near $0.50.
On placement, it matched GPT-5.6 Sol for third on Artificial Analysis and overtook Kimi K3. It is an Opus-class model tuned toward coding and agentic work, and xAI’s own documentation names it as the default model behind Grok Build.
That is a real product with real numbers, and it gets substantially less attention than the model that does not exist. Our AI models compared roundup places it against the current frontier.
The SpaceX Data Angle
The most interesting Grok 4.7 claim, and the one worth watching if the model does ship.
Musk has said xAI is adding SpaceX company data during supplemental training. That is unusual: a frontier lab training on proprietary engineering records from a sister company, in a domain where public training data is thin.
If accurate, it suggests a model with unusual depth in aerospace and hard engineering rather than one aiming at general benchmark leadership. That is a coherent strategy and a hard one to evaluate, because no public benchmark measures aerospace engineering competence.
It also raises questions nobody has answered. Whether that data creates capabilities that need disclosure, how it affects the model outside that domain, and what the arrangement means for two companies sharing a founder. xAI has not commented on any of it.
Worth holding lightly until there is documentation. This is a founder statement about training data, which is among the least verifiable claims a lab can make.
Why It Slipped, In Musk’s Own Words
The 11 September Grok 4.7 post is more informative than the four before it, because it names specific problems rather than giving a date.
Musk described the model as wrapping up prematurely on difficult tasks, and its answer-checking as not strict enough. Both are reinforcement learning tuning issues.
Those are real, recognisable problems. A model that gives up early on hard problems and does not verify its own answers rigorously is a model that will benchmark poorly on exactly the agentic and reasoning work Grok 4.7 is meant to lead. Fixing them is legitimate work.
They are also the kind of problems that take an unpredictable amount of time. “A few more days” for an RL tuning issue could be accurate or could be another month. The honest read is that a late-September launch is plausible and that any specific date should be treated as a guess until xAI publishes one.
How to Check for Yourself
Checking the Grok 4.7 status yourself beats every tracker, including this article, and takes thirty seconds.
Check xAI’s developer documentation at docs.x.ai. If grok-4.7 appears in the model list with a price and a context window, it has shipped. If the list ends at grok-4.6, it has not.
That single check is authoritative because a model that exists as a product must have an API identifier. No amount of founder commentary, tracker reporting or community speculation changes what is in the model directory.
Two secondary confirmations worth knowing: xAI’s release notes and its llms.txt file, both of which list current models. A genuine launch appears in all three roughly simultaneously.
What does not count as a launch: anonymous benchmark appearances, closed betas, invite-only access, A/B tests, and screenshots. Those have all preceded real launches before, and they have also preceded nothing at all.
What a Real Launch Looks Like
Useful context, because the Grok 4.7 situation is easy to mistake for normal pre-release quiet.
When GPT-6 Astra shipped on 3 September 2026, OpenAI published a model string, per-token pricing, a context window, a benchmark table and a system card the same day. When GLM-5.3-Flash shipped on 26 August, Zhipu put the weights on Hugging Face and the pricing in its documentation immediately. When Grok 4.6 itself shipped on 12 August, it arrived with a model ID and a price.
That is what a launch looks like across every lab, including xAI. The artifacts appear together, because a model that customers can call needs an identifier and a rate card to exist at all.
Grok 4.7 has produced none of those at any point across five announced windows. That is not a lab being quiet before a launch. It is the absence of a launch, and the distinction is checkable rather than a matter of interpretation.
If It Ships: What to Check First
Assuming a launch does arrive, the things worth verifying before believing any headline about it.
Does the parameter claim hold? 2.1 trillion against Grok 4.6’s 1.5 trillion is a founder figure. A model card either confirms it or it does not, and labs do not always publish parameter counts.
Is the token efficiency claim measurable? “Fewer output tokens per completed task” is testable the day an API identifier exists. Run the same prompts through Grok 4.6 and Grok 4.7 and compare total output tokens, not just the answer quality. That comparison is more useful than any launch benchmark.
How does the slower serving speed land in practice? Musk explicitly traded speed for capability. Whether that is acceptable depends on whether your work is latency-bound, which is the same question our AI inference API comparison works through for other providers.
Which benchmarks are cited, and at which version? Cross-vendor comparisons routinely quote different releases of the same test. That caveat applies to every model launch and it will apply to this one.
Does the SpaceX training show up anywhere measurable? If the supplemental data produces genuine aerospace capability, someone will find it. If nothing distinguishes Grok 4.7 in that domain, the claim was marketing.
Frequently Asked Questions About Grok 4.7
Has Grok 4.7 been released?
No. As of 12 September 2026, xAI’s model directory, release notes and documentation contain no grok-4.7 entry, no pricing and no context window. The newest available model is Grok 4.6.
When will Grok 4.7 be released?
Nobody has published a date, including xAI. Musk floated five windows between July and September 2026 and all passed. On 11 September he said the model needed “a few more days” of tuning without giving a new target.
What are the Grok 4.7 specifications?
Reportedly 2.1 trillion parameters against Grok 4.6’s 1.5 trillion, with better token efficiency and slightly slower serving speed. Every one of those figures is a Musk statement rather than an xAI documentation entry.
What is the newest Grok model I can use?
Grok 4.6, released 12 August 2026. It has a 500K context window, accepts text and images, costs roughly $2.00 input and $6.00 output per million tokens, and is the default model in Grok Build.
Is the SpaceX training data claim confirmed?
Only as a founder statement. xAI has published nothing about training data for this model, and claims about training corpora are among the hardest for anyone outside a lab to verify.
Why does the delay keep happening?
Musk’s most specific explanation, on 11 September, cited the model wrapping up prematurely on hard tasks and insufficiently strict answer-checking. Both are reinforcement learning problems with unpredictable timelines.
What This Says About AI Release Coverage
A broader point worth making, because Grok 4.7 is an unusually clear case of a general problem.
A large amount of AI coverage treats founder statements as equivalent to company announcements. They are not. A company announcement comes with a model card, a price, an API identifier and benchmarks that can be checked. A founder post comes with none of those and carries no commitment.
The gap shows up in predictable ways. Articles describing unreleased models in present tense. Comparison tables including specifications that exist only as claims. Release-date pages confidently stating windows that have already passed.
The practical habit that fixes this: before believing a model exists, look for its API identifier. Products have identifiers. Announcements have prices. Anything without both is a plan, however senior the person describing it.
That applies equally to the models this site covers favourably. Our GPT-6 Astra and GLM-5.3-Flash coverage cites model strings and published pricing precisely because those are the things that distinguish a shipped product from an intention.
Verdict
Grok 4.7 does not exist as a product you can use, and the confirmed record on it is remarkably thin for a model this widely discussed. Five release windows have passed, all of them from founder posts rather than company communication, and xAI has published no model ID, no price, no context window and no benchmarks at any point. The most recent signal, on 11 September, was an admission that the previous target had slipped alongside a description of unresolved reinforcement learning problems.
The honest caveat cuts the other way too. None of this means the model is vapour or that the delay is suspicious. Wrapping up prematurely on hard tasks and weak answer-checking are genuine problems worth fixing before release, and a lab taking longer to ship something that works is behaving correctly. A late-September launch is plausible. So is November.
Concrete next step: if you are waiting on this model, do not follow trackers or founder posts. Open docs.x.ai and look at the model list. When grok-4.7 appears there with a price beside it, it has shipped, and until then it has not, regardless of what any article including this one tells you.
Sources and Further Reading
- xAI developer documentation — the authoritative model list, and the only check that settles the question
- CellCog release-date analysis — documented the four-window timeline and the absence of xAI publications
- OrcaRouter Grok 4.7 tracker, the postponement reporting and what counts as a confirmed launch
- Julian Goldie’s confirmed-facts breakdown, Grok 4.6 specifications and the 11 September reinforcement learning statement
- TechJournal on the SpaceX data angle, supplemental training claims and earlier schedule changes
For related coverage on this site, see AI models compared for where Grok 4.6 sits against the current frontier, GPT-6 Astra for OpenAI’s shipped flagship, GLM-5.3-Flash for the open-weight alternative, and AI coding assistants for the tools these models compete inside.
[…] the 12 September target passed. Every one came from Musk personally rather than from xAI. Our full Grok 4.7 timeline separates what xAI has documented from what has only been […]