The rumor file closed on 21 September, the ordinary way: an API string appeared.
In August I kept two piles. One held what you could curl. The other held what Elon Musk had typed on X. The Grok 4.7 rumor file promised to move the lantern when an identifier showed up. It has. The product page is real. The docs answer to grok-4.7. The model card is a PDF with a date on it. Pretending the summer claims were never made would be a different kind of dodge. The honest move is to score them.
What held. The September window held, more or less. After 4.6 shipped in mid-August he had talked three to four weeks; the SpaceXAI note — SpaceXAI is the public name, a DBA of XAI LLC — landed on 21 September. Context stayed at half a million tokens. List price stayed where 4.6 sat: two dollars per million input tokens and six per million output when the prompt stays under two hundred thousand tokens. Cross that line and the whole request bills at the higher tier, four and twelve. Some gateways are discounting against that list this week; quote the SpaceXAI table first, then the promo. Function calling and structured outputs are on the card. Reasoning effort can be wound from low through medium and high to xhigh. You can buy it on the xAI API, in Cursor, and as the default in Grok Build. Gateways already list it. That is a shipped model.
The paperwork disagrees with itself on one small clock. The developer docs put a knowledge cut-off in May 2026. The model card talks June pretraining and supplemental data into August, including anonymized Cursor workflow traces. Prefer both lines over a single tidy date.
What did not hold. July’s size theatre did not make the card. “2.1 trillion parameters,” and the tidy story that 4.6 was 1.5T, still live in secondary write-ups. They do not live in the announcement, the docs, or the model card. Those pages say a larger base and longer reinforcement learning. They do not publish a parameter count. Treat the trillion as summer weather unless a primary source repeats it with a number.
The speed story flipped. In July the promise was “slightly slower to serve” in exchange for better work per token. The September line is the opposite theatre: same price and same speed as 4.6, with a separate fast variant sold at roughly twice the output speed and twice the price. Independent clocks put the main model somewhere near two hundred tokens a second and note that it burns more tokens per hard task than 4.6 did at xhigh. Faster serving and heavier spend can sit in the same product. The July bargain was not the one that shipped.
Where you can and cannot talk to it matters more than the scoreboard. On launch day it was alive for builders: Cursor, Grok Build, the API. The consumer surfaces — the website, the phone apps, Grok inside X — were explicitly deferred on the model card. Some secondary pieces already claimed it was live in the Grok app. Prefer the card. If someone tells you 4.7 is “on Grok” because they pay for Premium, ask which pane they mean. Premium is not a model card.
Cursor’s everyday fast model, Composer, was not renamed. The help text keeps it for cheap short work and points 4.7 at the long, hard jobs. The model card even admits the Cursor supplemental training. That is a coupling, not a costume change. SpaceXAI and Cursor are writing each other into the same release notes; they are not the same product.
The official scoreboard is loud and harness-shaped. Cursor’s own bench moves from about forty percent on 4.6 to the mid-forties on 4.7 at the highest effort. Terminal-Bench in the Grok Build harness nearly doubles. Artificial Analysis gives a small lift on its intelligence index and a clearer lift on its coding-agent index, with an Elo near sixteen fifty-seven on their Briefcase suite. Their terminal component sits softer than xAI’s harness number. Pick a number, name the effort and the harness, or you are quoting marketing.
None of this makes 4.7 the top of the class on aggregate indexes. The honest reading from the same day’s independent notes is improvement without coronation. That is still a release.
So the piles again, labeled. 4.6 remains the older sibling with a full public life. 4.7 is the new API citizen: same list price as its predecessor, live for tools and agents, not yet the voice in your X timeline, and still silent on the trillion-parameter boast that made July interesting. Grok 5 stays in its own drawer. Shipping a point release does not retire a year-end myth.
I said I would update when an identifier appeared. It appeared. The lantern is on grok-4.7 now, and the rumors that did not survive the card can finally stay in the file where they belong.