Musk: Grok 4.7 Comparable to Opus 5.0
Musk responded to a question from a user on social platform X, stating, "Grok 4.7 should be roughly comparable to Opus 5.0, not 5.1," adding that "it is stronger in some aspects and weaker in others," and mentioned that the team "needs to fix the multimodal performance."
He also announced the roadmap for subsequent versions—"Grok 4.8 will have significant improvements," "Grok 4.9 may reach Astra/Fable level," and "Grok 5 may surpass everything currently available," concluding with "we'll wait and see," indicating a noticeably more cautious tone than before.
This statement sharply contrasts with Musk's claim on September 2, when he publicly stated that "Grok 4.7 will surpass all existing models," and that "except for slightly slower service speed, it outperforms Grok 4.6 in all aspects"; however, by September 11, he admitted that the model "still needs a few more days of refinement," citing that the reinforcement learning training phase "may have overly penalized answer length," causing the model to give up too early on complex tasks and not fully self-validate.
On September 13, Musk skipped the originally planned Grok 4.7 and directly announced the larger parameter Grok 4.8 (approximately 2.5 trillion parameters, using a new C++ software stack), expanding nearly 20% from the initially planned 2.1 trillion parameters for Grok 4.7; as of now, xAI has not yet released the official product page, pricing, benchmark results, or specific release date for Grok 4.7 or 4.8.
According to previously disclosed timelines, the Grok 4.5, 4.6, and 4.7 models were originally planned for intensive release within about two months, with parameter sizes expanding from approximately 1.5 trillion for version 4.6 to 2.1 trillion for version 4.7; in comparison, Anthropic's recently released Claude Fable 5.1 achieved a score of 52.6% in the Terminal-Bench-Science benchmark, regarded as one of the current reference points in this field.
This cycle of "release—delay—retraction—further delay" occurs against the backdrop of xAI's ongoing investment in infrastructure—its Memphis "Colossus 2" supercomputing cluster has deployed approximately 550,000 GB200/GB300 chips, with a total power close to 1 gigawatt, and the GPU procurement cost alone is about $18 billion; for xAI, the gap between the speed of model iteration and actual performance is directly related to whether it can convince investors to continue funding this scale of computing power investment, also affecting the valuation narrative of Musk's enterprises in the AI lab ranking.
Source: Public Information
ABAB AI Insight
Musk's expectation management for the Grok series shows a clear pattern of "high-profile promises followed by gradual downgrades"—on September 2, he publicly stated that Grok 4.7 "will surpass all existing models," but by September 11, he admitted that there were flaws in the reinforcement learning training, and the model "still needs a few more days of refinement." On September 13, he simply skipped 4.7 and launched the larger parameter 4.8, now further lowering expectations to "roughly comparable to Opus 5.0, not 5.1"; this trajectory of "delayed release schedule and narrowed promotional rhetoric" is highly consistent with Musk's history of repeated delays and re-promising in projects like the Roadster hover demonstration and Full Self-Driving.
The frequent iterations are supported by xAI's ongoing investment in computing power—its Memphis "Colossus 2" supercomputing cluster has deployed about 550,000 NVIDIA GB200/GB300 chips, with total power close to 1 gigawatt, and the GPU procurement cost alone is about $18 billion. The logic behind this funding mobilization is to use extreme computing power to compensate for the flaws exposed in model training methods (such as excessive penalties on answer length in reinforcement learning), essentially using the speed of capital expenditure to hedge the risks of insufficient engineering maturity, attempting to continuously outpace competitors in parameter size, even if the actual performance of each model release does not meet expectations.
The handling of Grok 4.7's temporary delay due to training flaws and the direct launch of the next generation is similar to the occasional delays seen in model release schedules by OpenAI and Anthropic, but Musk's uniqueness lies in the fact that almost every time it is accompanied by highly publicized advance promotions and subsequent public self-corrections. This "promoting while correcting" approach is more akin to Tesla's historical expectation management of "fully autonomous driving is coming soon," rather than the relatively restrained release rhythm of traditional AI labs; the Grok series is currently widely regarded as lagging behind the latest flagship models from Anthropic and OpenAI, and the statement that Grok 4.9 "may reach Astra/Fable level" indirectly acknowledges this reality gap.
Essentially, this is a game about the rhythm of "technological substitution" in AI lab competition—when model training methodologies (such as reinforcement learning detail tuning) become bottlenecks limiting actual performance, simply piling on parameter size or computing power investment does not automatically translate into corresponding improvements in model capability. This is the fundamental mechanism behind the gap from "claiming to surpass everything" to "roughly catching up to half a generation earlier's Opus" for Grok 4.7; for the entire industry, this suggests that the model competition is gradually shifting from a rough stage of "competing on parameters and computing power" to a stage of "competing on the finesse of training methodologies," which is precisely a shortcoming that cannot be compensated solely by the speed of capital expenditure.
ABAB News · Cognitive Laws
The larger the parameters, the more it cannot cover the flaws in training methods.
First claiming to surpass everything, then saying to catch up to half a generation, in between is the real gap.
Computing power can buy scale, but cannot buy the finesse of training.