On September 13, 2026, Elon Musk announced on X that xAI's next-generation model, Grok 4.8, is completing its main training run, with 2.5 trillion parameters, built on a C++ software stack developed in-house by xAI from scratch, and that it will move straight into reinforcement learning once training finishes. At the moment of that announcement, Grok 4.7, with 2.1 trillion parameters, still had not been released — meaning xAI had announced a larger successor with a newer architecture before delivering its previous model.
Five Unfulfilled Promises
The delay history of Grok 4.7 is a winding one. According to a rundown by cellcog.ai, Musk gave at least five different release dates — from "within four weeks" on July 24, 2026, to "three to four weeks" on August 12, to "ten days" on September 1 — and every one of them slipped. On September 11 he offered his most concrete technical explanation to date: reinforcement learning training imposed an excessive penalty on the model's response length, causing the model to abandon difficult problems within its capability rather than persevering, and its self-checking was not strict enough.
Capability degradation during the RL alignment stage is not unusual for large models, but the case of Grok 4.7 points to a specific design trade-off: to make the model's output more concise, the training signal penalized overly long responses, and the side effect was weaker performance on high-difficulty tasks that require long chains of reasoning. This runs against the intuitive expectation that "more parameters means stronger."
On September 14, Musk further downgraded his public positioning of 4.7, describing it as "roughly comparable to Anthropic's Opus 5.0" rather than the across-the-board superiority to the strongest models on the market that he had previously implied.
The Architectural Bet Behind Grok 4.8
Grok 4.8's technical differences are not limited to the increase in parameters from 2.1 trillion to 2.5 trillion (a rise of roughly 19%). According to progressiverobot.com, xAI built an entirely new C++ training infrastructure for this version, running on 220,000 Nvidia GB300 GPUs interconnected by 800G high-speed network cards, with heavy use of pipeline parallelism. Musk claims this approach is more than "an order of magnitude" more training-efficient than mainstream JAX solutions.
Building your own training framework is a high-risk, high-reward choice. Mainstream frameworks such as PyTorch and JAX have accumulated years of community debugging and optimization; starting from scratch means xAI has to handle on its own all the low-level issues of numerical stability, the edge cases of distributed communication, and the details of memory management. Once it works, this stack will be tightly bound to xAI's own Colossus supercomputing cluster, and competitors cannot directly copy it. Grok 4.7 has already demonstrated the fragility of the RL stage; after switching to a self-developed training framework, Grok 4.8's risk points are no fewer.
The Moat of SpaceX Data
Grok 4.7's training incorporated SpaceX engineering data, including internal documents, telemetry, Starlink satellite records, and rocket development materials. According to bighatgroup.com, this supplementary training phase took place after the main pre-training ended and was the final step before release.
This is the only class of data asset xAI holds that its competitors are physically unable to replicate. OpenAI, Anthropic, and Google DeepMind all have no way to obtain SpaceX's engineering telemetry database.
Benchmark Coordinates: Starting from Grok 4.6's Actual Performance
To understand where 4.7 and 4.8 stand, one has to start from the only model of the previous generation that has been publicly released. According to benchmark data reported by finance.biggo.com, Grok 4.6 scored 61 on the AA Intelligence Index, level with GPT-5.6 Sol Max but below Claude Fable 5 Max's 62. On the GDPVal-AA v2 benchmark, Grok 4.6 scored 1753, leading same-generation competitors. But on Terminal-Bench v3.0, Grok 4.6 scored only 26%, clearly trailing GPT-5.6 Sol Max's 34.6%.
These numbers sketch a model with an uneven capability profile: comprehensive reasoning close to the top tier, but a structural shortfall in task types involving terminal operations and command-line execution. Grok 4.5 scored only 56 on the same index, and 4.6 rose to 61 — that generational jump is real. But Musk's positioning of 4.7 as "comparable to Opus 5.0" means the actual improvement of 4.7 over 4.6 is more than a step more conservative than the original marketing.
Pricing: The Only Undisputed Competitive Advantage
Against a backdrop of back-and-forth adjustments to performance positioning, the most quantifiable advantage of the Grok family today is price. According to progressiverobot.com, Grok 4.6's API pricing is $2 per million tokens for input and $6 for output, while competitors' flagship models typically sit in the $10/$50 range. That roughly fivefold price gap is enough to change how enterprises make their selection decisions in high-volume call scenarios.
If Grok 4.7 and 4.8 hold a similar pricing range at official release and can close 4.6's structural weakness on Terminal-Bench-type tasks, that will be the trigger point for real migration by enterprise users.
Roadmap and Signals to Watch
Musk has laid out a roadmap that is essentially a self-assessment: Grok 4.7 is roughly on par with Opus 5.0, Grok 4.8 is a clear improvement on 4.7, Grok 4.9 "probably reaches Astra/Fable level," and Grok 5 is positioned as xAI's first AGI-level model. This roadmap is not based on a published benchmark card, nor has it been independently confirmed by any outside party — it is the founder's self-description on social media.
The experience of Grok 4.7 adds a baseline footnote to that roadmap: there is a structural time lag between announcement and delivery, and performance expectations undergo downward revision as release approaches. This does not mean the roadmap itself is wrong, but developers who rely on API stability for product planning need to keep enough of a buffer.
Three signals will serve as concrete observation points for gauging the roadmap's credibility: Grok 4.8's score on Terminal-Bench v3.0-type tasks at official release; whether the official release comes with a complete benchmark card, API documentation, and pricing table; and the actual time span between the completion of RL training and the official launch. If all three are in place by mid-October 2026 and the Terminal-Bench score improves markedly, xAI's roadmap will have substantive support.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接