[Digital Today intern reporter Seung-a Yoo] Elon Musk (일론 머스크) said the performance of xAI's delayed artificial intelligence (AI) model Grok 4.7 will be similar to Anthropic's Opus 5. He also said it still needs improvement compared with Opus 5.1 or OpenAI's Astra.

Cryptopolitan, a blockchain outlet, reported on Sept. 14 that Musk wrote in a post on X, formerly Twitter, that Grok 4.7 "should be roughly on par with Opus 5" and that it is "better in some ways, worse in others". He acknowledged that multimodal performance also needs improvement.

He also explained the reason for the release delay. Musk said Grok 4.7 "needs a few more days of polishing" and that it may have given an excessively large penalty to response length during the reinforcement learning process. This led to a problem in which the model "gives up too early" even on difficult tasks it can actually solve, and he said its ability to re-check its own results is still not sufficiently rigorous.

Grok 4.7 was initially slated for release on Sept. 12, but its launch was postponed. What has been known so far is that it is a model with about 2.1 trillion parameters, and it was reported that it was reinforced with SpaceX engineering data. If true, it would be about 40 percent larger than Grok 4.6, which is known to have about 1.5 trillion parameters. Grok 4.7's official benchmark scores and token pricing have not yet been confirmed. A claim that it could be up to 10 times cheaper than rival models also remains unverified.

Anthropic's Claude Fable 5.1, cited as a comparison, has already been released, with external assessments and pricing information available. Fable 5.1 scored 52.6 percent on Terminal-Bench-Science and is offered at $10 per 1 million input tokens and $50 per 1 million output tokens.

Musk also presented a roadmap for follow-up models after Grok 4.7. He said Grok 4.8 is a 2.5 trillion-parameter model being trained using xAI's new C++ software stack, and that it will finish training within this week before moving into reinforcement learning. He said Grok 4.8 will show a "noticeable improvement" over Grok 4.7.

He claimed Grok 4.9 may reach a level similar to OpenAI's Astra and Anthropic's Fable series. He also said the 2.5 trillion-parameter model performs better than the existing 2.1 trillion-parameter model, and explained that some errors were fixed midstream during training with JAX. He then signalled plans to train a 3 trillion-parameter model using improved internal training software and refined data.

Grok 5 is the model he is most excited about. Musk said Grok 5 "maybe better than anything" and that results will have to be watched. Asked which model can implement a specific capability, he replied, "That's what Grok 5 will do." He did not provide a release timing for Grok 5 or specific evidence to support its performance.

xAI's baseline model that can currently be verified in the market is Grok 4.6. Opus 5.1 and OpenAI's Astra, which Musk cited as performance targets, are already released models, while Grok 4.7 is undergoing additional tuning ahead of release. As a result, Grok 4.7's actual competitiveness is expected to be confirmed after its official launch and independent benchmark results are released.

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.