Grok has been the topic of a lot of Elon Musk’s attention lately, especially given the fact that he recently mentioned that 4.7 needs a few more days to cook. With that in mind, he’s posted a near-term layout of what to expect from SpaceXAI in terms of Grok development, especially with how Grok will compare against peer models from Anthropic and OpenAI.
Musk confirmed that Grok 4.7 should be competitive with Opus 5.0, not Anthropic’s latest Opus 5.1 build, at least in some capabilities. One of the key things holding it back have been multimodal capabilities, and the training team seems to be hard at work on that.
However, he also spoke about what’s coming further down the ladder.
Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance.
Grok 4.8 will be a noticeable improvement.
Grok 4.9 is probably Astra/Fable class.
Grok 5 maybe better than anything. We shall see.
— Elon Musk (@elonmusk) September 14, 2026Grok 4.8 will be a more noticeable improvement over 4.7, likely being faster and smarter, and passing more key coding and intelligence testing than 4.6 does today and 4.7 will do soon. However, Grok 4.9 is intended to meet or match OpenAI’s recently released Astra and Anthropic’s Fable models in intelligence, which will be a massive capability leap, especially for users of Grok Build or Grok Bot.
With OpenAI claiming they’re getting closer to AGI than ever with the release of Astra, that means Grok 5 will likely also be quite close to SpaceXAI’s goals of building an AI that can help answer the questions of the universe.
Grok 4.7: Close, but not Quite
Musk’s own confirmation that 4.7 is a step forward over current builds of Grok, but not comparable to leading frontier models is key here. It means that the team at SpaceX knows they’re still in for the long haul for model improvement.
Bringing Grok to roughly Opus 5.0 parity is still a major step forward, as xAI was only stood up 3 years ago, with Grok’s 3-year anniversary coming up this November. Anthropic’s first model was released in March of 2023, meaning that they have roughly a six month lead on SpaceX, with that gap narrowing with every new release.
Multimodal support for a model might not be the thing that everyone focuses on, but having a flagship model being able to work on more than just code is important. Grok serves as the same core that runs Grok Voice models, Grok Chat Agents, Grok Imagine, Grok Bot, Grok Build, and the Grok on the web and on phone apps.
Each of those services uses a different modality, and many users often use Imagine to generate images or even the Grok app’s ability to use a camera to guide them through a particular real-life task. Flagship models absolutely need multimodal support, and with SpaceX focusing on catching up here, Grok is likely going to become even more useful for the average user with 4.7 and 4.8.
Grok 4.8 Begins Reinforcement Learning
Musk didn’t mention exactly what Grok 4.8 would be comparable to in his posts, but called it a notable improvement over 4.7. He also went on today to confirm that 4.8 will be finishing the final training steps this week, and begin the next and final phase – reinforcement learning.
Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL
— Elon Musk (@elonmusk) September 14, 2026With Grok 4.8 already entering the RL phase, you might ask why SpaceX is even bothering to release 4.7. The fact is, with each iterative release, the team can gather more training data and better refine the models in their training phases, before they even enter reinforcement learning.
That can help make larger leaps in capability than withholding a model for months. In addition, Musk also confirmed that Grok 4.8 seems to be a 2.5T model, and 4.9 is likely a 3T model, with each leap in parameters a functional leap in capabilities.
It’s unequivocally better than the 2.1T that was trained with Jax, but we made some mistakes that were only corrected mid run and we have greatly improved the data quality in the past few months.
The upcoming 3T run with improved internal training software and much better data…
— Elon Musk (@elonmusk) September 14, 2026The Big C++ Rewrite
Aaron Burnett, part of the the Cursor and xAI teams working on Grok Bot, also chimed in with some additional details on exactly why the team has been able to move so fast in the background. The SpaceXAI team was focused on rewriting the software they use to train Grok itself, which was previously done with Jax.
Now, the new training suite uses C++, which has cut the reinforcement learning time for models in half for the ~2T parameter models, and makes substantial time savings for even larger models.
After the C+++ rewrite, our modeling indicated RL time for these models starting at 4.8 COULD be cut in half for ~2T models and substantially more time savings on a % basis for larger models. The thing I’d suggest watching is time between releases once they get some of the kinks… https://t.co/NHts78gaIe
— Aaron Burnett (@aaronburnett) September 14, 2026With the larger 2.5T run on the new C++ stack wrapping up training this week, that means that Grok 4.8 could be a closer release than we expect, following hot on the tails of Grok 4.7. But it also means that 4.8 is really just an engineering step forward, and what we really should expect from Grok 4.7 and 4.9 in the future.
Grok 4.9 Should Be A Fable Class Model
Grok 4.9 is supposed to be the model that finally brings SpaceXAI on par with its two key competitors, OpenAI and Anthropic. Meeting the Astra/Fable class will truly make Grok a frontier AI model, and one that people will legitimately want to select over other competitors today.
For now though, we’re still expecting Grok 4.9 to launch sometime this year, but its likely many weeks away at best, if not closer towards December. Each newer and larger model takes more time to train, even with significant improvements like the new C++ training pipeline.
With that all said, there’s a lot to look forward to in the coming weeks – with 4.7’s launch around the corner, hopefully as soon as this weekend.
