THE LADDER, IN MUSK'S OWN WORDS
● Grok 4.7 — roughly on par with Opus 5.0
● Grok 4.8 — a noticeable improvement on 4.7
● Grok 4.9 — Astra or Fable class
● Grok 5 — may be better than anything, and where he expects AGI
● Available today: none of them.
The useful thing about this roadmap
Musk benchmarked each unreleased version against a model that has already shipped. That is unusually easy to work with, because it means every rung on the ladder has a real product you can use right now at the claimed level.
So the question stops being "when does Grok 5 arrive" and becomes "what already does the thing Grok 5 is promised to do".
Each rung, and what matches it today
| Promised |
Claimed level |
Shipped equivalent |
| Grok 4.7 | Opus 5.0 class | Claude Opus 5 — available on Claude Max and via API |
| Grok 4.8 | Above 4.7 | Nothing named. The gap between Opus 5 and Astra |
| Grok 4.9 | Astra or Fable class | GPT-6 Astra on ChatGPT Plus · Claude Fable 5.1 at $10/$50 |
| Grok 5 | Better than anything | By definition, nothing yet |
THE ROW THAT MATTERS IS THE THIRD ONE
Grok 4.9 is promised at Astra or Fable class. Astra shipped on 3 September and is included in ChatGPT Plus at $20. Fable 5.1 is generally available at $10 in and $50 out.
Both have published prices, model cards and APIs. Grok 4.9 has a sentence on X. If the capability is what you need, it exists today and you can buy it.
What is actually known about each
Grok 4.7
Roughly 2.1 trillion parameters with SpaceX engineering data in supplemental training, per Musk. Scheduled for 12 September and still unreleased. Delayed because reinforcement learning may have penalised response length too heavily, leaving the model prone to giving up on hard tasks early.
Grok 4.8
2.5 trillion parameters on xAI's new C++ training stack. Musk said on 14 September it would finish training that week and then start reinforcement learning. Which means RL — the stage that delayed 4.7 — has not begun.
Grok 4.9
A claimed capability level and nothing else. No parameter count, no training detail, no date.
Grok 5
Asked how close 4.8 would come to AGI, Musk replied "That will be Grok 5". There is no agreed definition of AGI, no benchmark that settles it, and no date.
How Musk's dates have gone
| Target given | What happened |
|---|
| Grok 4.7, late August | Missed |
| Grok 4.7, first week of September | Missed |
| Grok 4.7, 12 September | Missed. Still unreleased |
That is not a criticism of the engineering, which nobody outside xAI can assess. It is a reason to plan around shipped models rather than announced ones.
What to do while you wait
- If you need Opus 5 class today, it is on Claude Max and the API. That is the 4.7 level, available now.
- If you need Astra or Fable class, both shipped this month with published pricing. That is the 4.9 level, available now.
- If you are on SuperGrok, model upgrades have historically reached every paid tier including the $10 Lite plan, so the version is not a reason to change tier.
- If you are building on the API, nothing changes until a model ID appears at docs.x.ai. There is none for 4.7, 4.8, 4.9 or 5.
- If you want the live X data that no competitor has, that is available on the current Grok and does not depend on any of this.
Sources
FAQ
When is Grok 5 coming out?
No date has been given. Musk named it as the version where he expects AGI, with no timeline attached, and xAI has published nothing about it.
What will Grok 4.9 be as good as?
Musk described it as Astra or Fable class. Both of those models shipped this month and are available today with published pricing.
Is Grok 4.7 better than Opus 5?
Musk described them as roughly on par. Grok 4.7 is unreleased, so the claim cannot be tested. Opus 5 is available on Claude Max and via the API.
When is Grok 4.8 released?
No date. As of 14 September it was described as finishing training that week, with reinforcement learning still to come — and reinforcement learning is what delayed 4.7.
Will Grok 5 be AGI?
Musk said so. There is no agreed definition of AGI and no benchmark that resolves it, so the claim is not currently testable by anyone.
Should I wait for a newer Grok before subscribing?
Model upgrades have reached all paid Grok tiers rather than the top one only, so the version is not the variable. Decide on the usage headroom you need now.