Musk admits Grok 4.7 is behind Anthropic, OpenAI's latest models; promises AGI by Grok 5

Elon Musk has tempered expectations for xAI’s delayed Grok 4.7 model even before it is released, conceding in X posts that the model lines up with Anthropic’s Opus 5.0 more favorably than in comparison with the newer Opus 5.1 model, Fable or even Astra from OpenAI. 

Musk also said that multimodal performance still needs work. The admissions, along with the performance, are unlike the hype the Tesla CEO typically stokes before products reach the market. 

What did Musk say about Grok 4.7? 

Responding to queries about how Grok 4.7 stacks up against the competition, Musk wrote that it “should be roughly on par with Opus 5.0, not 5.1.” The SpaceX chief executive added that it was “better in some ways, worse in others.” 

Musk explained on Friday that Grok 4.7 needed “a few more days to cook,” because apparently, the team may have penalized response length too heavily during reinforcement learning. As a result, the new model “gives up on hard tasks (that it can do!) too early,” and its ability to double-check its own work was not yet at the required level.

Grok 4.7, originally scheduled for a September 12 release date, has been described as a roughly 2.1-trillion-parameter model supplemented with SpaceX engineering data. If those reports are confirmed, the latest model would be 40% larger than Grok 4.6, which is reported at about 1.5 trillion parameters. 

Third-party assessments of Anthropic’s Claude Fable 5.1 already report scores of 52.6% on Terminal-Bench-Science, and a published price of $10 per million input tokens and $50 per million output tokens. 

The unreleased Grok 4.7 has no confirmed benchmark scores or token pricing. As of this report, the claim that it could be up to 10 times cheaper than a competing model is still unverified.

Musk is already promising AGI 

Musk did not take long to return to his ambitious posting, sketching a rapid release sequence to follow the still unreleased 4.7 model. 

Grok 4.8, he said, is a 2.5-trillion-parameter model trained on xAI’s new C++ software stack that will finish training this week and then begin reinforcement learning. He called it “a noticeable improvement” over 4.7.

From there, Musk claimed Grok 4.9 would probably be “Astra/Fable class,” a reference to OpenAI’s GPT-6 Astra and Anthropic’s Fable line. He described the current 2.5T model as better than an earlier 2.1T version trained with Jax that suffered mistakes “only corrected mid run,” and pointed to a coming 3-trillion-parameter run with upgraded internal training software and cleaner data.

The bigger claim was reserved for the top of the ladder. Musk said Grok 5 “maybe better than anything,” hedging with “we shall see.” In a separate reply to a user, asked what would deliver a specific capability, he answered only: “That will be Grok 5.”

Musk offered no date for that release and no evidence beyond the assertion. For now, the measurable baseline remains Grok 4.6, and the models Musk himself named as the bar to clear, Opus 5.1 and OpenAI’s Astra, are already in the market while Grok 4.7 waits to launch.

The smartest crypto minds already read our newsletter. Want in? Join them.