Grok 4.7 has turned up on a Google Cloud quotas page, and that is the strongest pre-release signal the model has produced so far. The sighting was reported on 17 September 2026 at 07:10 UTC by the AI watcher Chris, posting as @ChrisGPT, who tied it to an earlier identifier found ten days previously and concluded that a launch is “presumably imminent.”
Neither xAI nor Google has announced anything. As of this writing, xAI’s own model catalogue still tops out at grok-4.6, Google’s Vertex AI partner-model listing still tops out at Grok 4.6, and there is no Grok 4.7 model card, price or release note anywhere in first-party documentation. What exists is plumbing: a configuration identifier, and now a quota entry.
That is worth taking seriously and worth taking carefully. Quota entries have a good record of preceding launches by hours rather than weeks, but this particular model has already slipped four publicly stated windows. This article sets out exactly what was found, why a quota page is a meaningful signal, what the first-party sources currently say, what the prediction markets are pricing at the time of writing, and how the sighting compares with a list of missed deadlines that is now two months long. Our earlier piece on Gemini 3.8 Live slugs on the same cloud quota page covers the parallel Google sighting.
Table of contents
- What Was Actually Spotted and Where
- Why a Quota Entry Is a Real Grok 4.7 Signal
- What xAI’s Own Catalogue Says Right Now
- What Musk Has Said About Grok 4.7
- What the Prediction Markets Price a Grok 4.7 Release At
- A Record of Missed Grok 4.7 Windows
- How to Read Pre-Release Plumbing Signals
- Frequently Asked Questions About Grok 4.7
- References
What Was Actually Spotted and Where
The claim is specific, and it is worth separating the two artifacts it rests on.
| Date | Artifact | Where it surfaced | What it establishes |
|---|---|---|---|
| 7 September 2026 | Identifier grok-4-7-0907 | Cursor support, via a configuration change in Grok Bot | A build exists and has a dated slug |
| 17 September 2026 | Quota entry for grok-4.7 | Google Cloud quotas page | A second platform is provisioning capacity for it |
The post, in full
“The first Grok 4.7 identifier we found came from Cursor support on September 7, when a configuration change exposed grok-4-7-0907 in Grok Bot. Now there’s a reported Google Cloud quota entry for grok-4.7, suggesting another integration is being prepared. That makes me think we could see it tomorrow or sometime next week. Still speculation date wise, – but presumably imminent.”
The two artifacts prove different things
The Cursor identifier shows that a build of Grok 4.7 has a name and was reachable enough from a coding product’s configuration to leak. The quota entry shows something further along: a distribution partner has created the capacity accounting a model needs before it can be served. Neither is an announcement, but the second is closer to one.
The hedge in the source is doing real work
Chris explicitly labelled the date “speculation.” That distinction matters because the artifact is verifiable and the timing inference is not. A quota row is a fact about a console; “tomorrow or next week” is a guess about a launch calendar that has already moved repeatedly.
The reach was modest
The post carried 9,372 views, 97 likes and 10 bookmarks at the time of reading — small numbers by the standards of this beat, and a reminder that this is a watcher signal rather than a broadly reported story.
Why a Quota Entry Is a Real Grok 4.7 Signal
Cloud quota pages are not marketing surfaces. They exist because capacity has to be metered, and the metering has to be defined before anyone can call the model.
Partner models get provisioned like first-party ones
Grok is available on Google Cloud as a partner model, alongside the platform’s own families. Serving any of them requires per-project quotas keyed to a base model identifier, which is why a new model name appears in the quota tables as part of the plumbing rather than as part of the launch.
The entry precedes the documentation
In the sequence Google follows, quota definitions and console entries tend to land before the public model card and pricing page. That ordering is what makes the page a leading indicator: the console reflects what the platform is configured to serve, and the docs reflect what the platform is ready to announce.
The observed lag has been short
Watchers tracking this pattern have generally seen a quota appearance resolve into a release the same day, or within about twelve hours. That is the basis for the “imminent” read, and it is the single strongest argument for the sighting mattering.
But short lags are a tendency, not a rule
A quota entry can also be staged days in advance, or created for a limited regional preview, or left in place after a slipped date. The Gemini 3.8 Live slugs spotted on the same console two days earlier had not produced a public launch either. Treat the signal as evidence that Grok 4.7 is being prepared for serving, not as a date.
What xAI's Own Catalogue Says Right Now
The first-party position is unambiguous and unchanged: there is no Grok 4.7.
| Model | Context | Input per 1M | Output per 1M |
|---|---|---|---|
| grok-4.6 | 500k | $2 to $4 | $6 to $12 |
| grok-4.5 | 500k | $2 to $4 | $6 to $12 |
| grok-4.3 | 1M | $1.25 to $2.50 | $2.50 to $5 |
| grok-4.20-0309-reasoning | 1M | $1.25 to $2.50 | $2.50 to $5 |
| grok-4.20-multi-agent-0309 | 1M | $1.25 to $2.50 | $2.50 to $5 |
| grok-build-0.1 | 256k | $1 to $2 | $2 to $4 |
No model page, no price, no release note
The developer catalogue lists grok-4.6 as the newest text model and carries no Grok 4.7 entry of any kind. Grok 4.6 was released on 12 August 2026 as the long-running-agent model, which means the current generation has been the newest for over a month.
Vertex lists one version behind as well
Google Cloud’s own partner-model documentation for xAI currently names Grok 4.1 Fast, Grok 4.20, Grok 4.3 and Grok 4.6. A quota entry for grok-4.7 sitting alongside a documentation set that stops at 4.6 is exactly the shape a pre-launch staging step takes — and exactly the shape a stale row would take too.
Pricing gives a floor for expectations
If Grok 4.7 follows the 4.5 and 4.6 pattern, it lands at $2 to $4 per million input tokens and $6 to $12 per million output with a 500k context window. The older 4.20 and 4.3 lines are cheaper and offer a 1M window, so a 4.7 at 4.6 pricing would continue the pattern of paying a premium for the newest reasoning line rather than for context length.
What Musk Has Said About Grok 4.7
The gap between the plumbing and the product is best understood through the public statements, which are unusually specific and have repeatedly overshot.
| Date | Statement | Implied date |
|---|---|---|
| 24 July 2026 | “4.6 in 2 weeks and 4.7 in 4 weeks” | About 21 August |
| 28 July 2026 | 4.6 “around August 7” at 1.5T; 4.7 at 2.1T parameters | Not dated |
| 12 August 2026 | “significantly better than 4.6 and should be ready in 3 to 4 weeks” | 2 to 9 September |
| 2 September 2026 | “Grok 4.7 comes out in 10 days” | 12 September |
| 11 September 2026 | “needs a few more days to cook” | Mid-September |
| 14 September 2026 | “roughly on par with Opus 5.0, not 5.1… We need to fix multimodal performance” | Not dated |
The slip, in days
The reason given for the delay is a training problem
On 11 September, Musk said Grok 4.7 “needs a few more days to cook,” explaining that “we might have penalized response length too much (or something) in RL, as it still gives up on hard tasks (that it can do!) too early and isn’t yet sufficiently rigorous in checking its work.” That is a reward-shaping issue in reinforcement learning, not a packaging delay — which is consistent with a model that is being retrained rather than one that is waiting on a launch slot.
The expectation has been revised downward
Three days later, the framing shifted from “exceed all current models” to “roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance.” Anthropic’s published catalogue lists Fable 5.1, Opus 5, Sonnet 5 and Haiku 4.5, with no Opus 5.1, so the comparison is against the top of a line rather than against a specific rival release.
The successor is already in training
The same week, Musk said Grok 4.8 — described as a 2.5T-parameter model trained on a new C++ software stack — would finish training and begin reinforcement learning. We covered that claim in detail in our piece on Grok 4.8’s 2.5T model and C++ stack. A successor entering RL while its predecessor is still unreleased is the clearest sign that the Grok 4.7 delay is about quality gating rather than capacity.
What the Prediction Markets Price a Grok 4.7 Release At
Two venues carry live contracts, and both moved up on the day of the quota sighting.
Kalshi moved on the day
The Kalshi contract on a release before 18 September last traded at 64 cents, against a previous price of 63, with a bid of 67 and an ask of 72 — a spread sitting above the last trade, which is what an order book looks like when demand is arriving. Open interest was about 1,216 contracts on roughly 2,077 of volume, with 245 in the previous 24 hours. The 25 September contract last traded at 72 cents against a previous 68.
Polymarket is more confident and better capitalised
Polymarket’s contract on the next Grok model arriving by 18 September traded at 76 cents with a 74 bid and a 78 ask, on about $21,988 of volume within an event carrying roughly $39,769. The tighter spread and larger book make it the more informative of the two numbers.
The day-by-day ladder is far less certain than the aggregate
Polymarket also runs a per-date market, and it disagrees with its own weekly contract in an instructive way.
Read the thin markets as mood, not forecast
The per-date contracts carry a few hundred dollars of volume each and, in several cases, bids at a fraction of a cent against asks near 40 cents. Those spreads make the mid-prices close to meaningless as forecasts. The weekly contracts, with real two-sided books, are the ones worth quoting.
Grok 5 is priced at essentially zero
The Kalshi contract on a Grok 5 release before 1 October last traded at 1 cent with about 13,678 contracts of open interest — a large, settled book that has concluded the next release will carry a 4-series number. That is a useful cross-check on the Grok 4.7 read rather than a separate story.
A Record of Missed Grok 4.7 Windows
The most sobering context for the quota sighting is the resolution history of the contracts that came before it.
Seven consecutive deadlines resolved No
Kalshi has settled seven Grok 4.7 release deadlines during August and September 2026 — before 14, 21, 28 and 31 August, and before 4, 8 and 11 September. Every one resolved No. The equivalent Grok 4.6 series resolved No five times before settling Yes on the 13 and 15 August contracts.
The 4.6 precedent is mildly encouraging
Grok 4.6 also slipped repeatedly before arriving, resolving No on contracts dated 6, 7, 8, 9 and 10 August before resolving Yes on the 13 August deadline. A model that misses several windows and then ships is the house pattern here, not an anomaly.
But the cause of the delay is different this time
Grok 4.6’s slippage read as scheduling. The 11 September explanation for Grok 4.7 named a specific reinforcement-learning defect: a model that gives up early on tasks it can do, and does not check its work rigorously enough. Fixing a reward-shaping problem is an open-ended amount of work, and it is not obvious that a quota entry created before that fix lands means the fix has landed.
Two signals now point in opposite directions
The quota entry says the serving path is being prepared. The training commentary says the model was not behaving correctly six days ago. Both can be true at once — platforms provision ahead of readiness — which is why the honest read of the Grok 4.7 sighting is “soon and not certain” rather than “tomorrow.”
How to Read Pre-Release Plumbing Signals
The Grok 4.7 case is a good template for weighing this class of evidence generally.
| Signal | Strength | Typical lead time |
|---|---|---|
| First-party model card or price page | Conclusive | Zero — it is the launch |
| Cloud quota or console entry | Strong | Hours to a few days |
| Partner documentation update | Strong | Same day to days |
| Dated identifier in a third-party product config | Moderate | Weeks, and can be a build that never ships |
| Executive statement of a timeframe | Weak on timing | Historically overshot four times here |
| Prediction-market price | Useful only where the book is deep | Reflects the above, with a lag |
Identifiers age badly
A dated slug such as grok-4-7-0907 records when a build was cut, not when it will ship. Ten days after that identifier appeared, the model was still being retrained. Treat a leaked identifier as proof of existence and nothing more.
Quota entries are the best early signal available
Of everything short of an announcement, a quota row is the artifact most tightly coupled to actually serving traffic. That is why the 17 September sighting moved both markets on the day, and why it deserves the attention it got.
Executive timeframes are the weakest input
Four stated Grok 4.7 windows have passed. A forecaster who had weighted the statements heavily would have been wrong four times; one who had waited for platform artifacts would have had nothing to act on until 7 September and a real signal on 17 September.
What would confirm it
A model card at the xAI developer documentation, a Grok 4.7 page in Google’s partner-model listing, or a price row in the catalogue. Any one of those settles it. Until then, the honest summary is that Grok 4.7 exists, is being provisioned, and does not yet have a public release.
Frequently Asked Questions About Grok 4.7
Has Grok 4.7 been released?
Not at the time of writing. There is no model page, price or release note in xAI’s documentation, and Google’s partner-model listing still tops out at Grok 4.6.
How big is Grok 4.7 expected to be?
Musk said on 28 July 2026 that Grok 4.7 would be a 2.1T-parameter model, against 1.5T for Grok 4.6.
How good is it expected to be?
The most recent public framing, on 14 September, is “roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others,” with multimodal performance named as needing work.
Why has it been delayed?
The stated reason is a reinforcement-learning issue: response length may have been penalised too heavily, leaving the model giving up early on tasks it can complete and not checking its work rigorously enough.
What is the quota page and why does it matter?
Google Cloud meters partner models such as Grok through per-project quotas keyed to a base model identifier. A new name appearing there means capacity is being configured, which usually happens shortly before a model can be called.
Where does Grok 4.8 fit?
Musk has described Grok 4.8 as a 2.5T-parameter model trained with a new C++ software stack, which was due to finish training and begin reinforcement learning in mid-September — while Grok 4.7 was still unreleased.
Readers tracking releases across vendors may want our AI models and tools hub, and teams deciding which model to standardise on can start with our AI strategy work.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.