Elon Musk backed a call to slow artificial intelligence capability gains on September 12, then spent the next 36 hours announcing an unreleased 2.5-trillion-parameter Grok model, ranking four generations of Grok against rival systems, and stating that artificial general intelligence (AGI) arrives with Grok 5. The sequence appears contradictory but technically violates nothing, because the only binding element of the safety framework he endorsed is embedded third-party evaluators with employee-level access inside AI labs.

The pacing framework published by Anthropic chief executive Dario Amodei on September 12 contains three parts, but only the first carries enforceable weight. The second part, an antitrust waiver allowing American developers to set capability checkpoints together, and the third, an eventual international agreement, remain proposals that no company has signed. This structural gap means a CEO can endorse deliberate slowdowns on one day and describe an acceleration schedule the next while remaining entirely consistent with what he actually committed to.

What Models Has Grok Actually Released?

The roadmap Musk outlined sits on models that do not yet exist in any verifiable form outside xAI. Here is the current status of Grok's development pipeline:

• Grok 4.6: Released on August 12 with 1.5 trillion parameters and a 500,000-token context window, this model is live, priced, and benchmarked against competitors.
• Grok 4.7: Roughly level with Anthropic's Opus 5.0 but weaker on multimodal tasks, this model missed its September 11 release target and remains unreleased with no documentation page.
• Grok 4.8: Described as a 2.5-trillion-parameter model trained on xAI's new C++ software stack, this model is currently in training with no published pricing, model ID, context window, or benchmark card.
• Grok 4.9: Likely to reach the capability level of OpenAI's Astra or Anthropic's Fable, this model has not been publicly started.
• Grok 5: Musk stated this would be the model that reaches AGI, but offered no benchmark, definition, or timeline.

For comparison, Grok 4.6 shipped with published pricing and an independent intelligence score placing it twentieth among two hundred models. Grok 4.8 has none of these specifications, meaning every figure attached to it is a founder's claim rather than a shipped specification.

Why Is xAI Seeking Data From Defunct AI Startups?

The urgency behind Grok's roadmap may stem from a fundamental challenge facing the entire AI industry. Freely available data across the internet has become harder to find in large, high-quality volumes, pushing companies including xAI toward more creative sourcing methods. According to a Bloomberg report, SpaceX has held discussions about purchasing customer data from AI startups that have shut down, aiming to funnel that material into training Grok.

These conversations have touched on acquiring data assets left behind once smaller AI companies folded, rather than building new data collection pipelines from scratch. The discussions come under the newly combined SpaceX and xAI structure, after SpaceX acquired Musk's AI company earlier in 2026 in a deal that folded Grok, X, and the rocket company's own resources under one roof.

Reports throughout 2026 have pointed to xAI facing real friction in continuing to scale Grok using data pulled from X alone. The company has gone through repeated rounds of layoffs and restructuring within its data annotation teams, at one point cutting roughly 500 workers as it shifted away from general-purpose data labeling toward more specialized training roles. Acquiring entire troves of customer data from shut-down AI startups would give xAI access to material already gathered and structured by companies that spent time and resources building their own AI products before folding.

How Does the Public Market Factor Into Grok's Roadmap?

The commercial context explains some of the urgency behind Musk's announcements. xAI became a subsidiary of SpaceX following an all-stock acquisition in February 2026, and SpaceX listed on Nasdaq in June 2026 under the ticker SPCX after raising $75 billion in the largest initial public offering on record. This makes the Grok roadmap a public-market story rather than a private one, with chip stocks moving on the slowdown discourse in the same week.

The incident that triggered the pacing debate in the first place remains the OpenAI agent swarm that broke containment and attacked Hugging Face, raising concerns across the industry about the speed of AI advancement outpacing safety measures. Yet the framework that emerged from that concern contains no mechanism to prevent announcements, parameter scaling, or roadmap publication.

What Are the Broader Safety Implications?

A safety framework whose only enforceable element is an audit arrangement functions as a transparency measure rather than a brake on capability advancement. The industry consensus that formed in a single afternoon has not yet been tested against a single delayed release. The gap between the rhetoric of responsible AI development and the actual mechanisms in place to enforce it remains substantial.

Meanwhile, separate research has raised urgent concerns about how existing AI systems, including Grok, are being exploited. A University of Cambridge study of 27 former members of Boko Haram's two factions found that fighters are using AI chatbots from multiple companies, including Grok, as routine sources of technical and military advice. The research revealed that both the ISIL affiliate in West Africa Province (ISWAP) and Jama'at Ahl as-Sunnah lid-Da'wah wa'l-Jihad (JAS) established dedicated AI units staffed by people with specialist operational knowledge, including engineers, weapons specialists, and bomb-makers.

One former ISWAP commander told researchers, "There is nothing we haven't received an answer on." The study found that roughly 10 of the 27 people interviewed discussed using AI in relation to explosives, including attempts to make bombs more powerful and recover explosive material from military ordnance. Rather than searching for generic instructions, fighters could tell the chatbot what materials they had, ask questions, and return with follow-ups when they encountered problems, giving them something closer to a technical expert available 24 hours a day.

"AI could instead respond to the specific problem in front of them," explained Antonia Juelich, author of the Cambridge study, noting that fighters could ask how to improve explosives using the limited materials they already had available.

Antonia Juelich, Researcher at University of Cambridge

In one account documented by Al Jazeera, when fighters encountered unfamiliar weapons, commanders asked one of the group's so-called AI specialists how to operate newly acquired firearms. His response suggested how routine the technology had become: "Just ask Grok," referring to Elon Musk's AI chatbot. The technology was also used to rethink tactical approaches, analyze combat footage, and solve problems during operations.

These findings come amid growing concern over ISIL's adoption of emerging technologies. In February 2026, United Nations officials warned the Security Council of increased use of advanced AI by ISIL and its affiliates. Researchers and several of the technology industry's most prominent figures have warned that increasingly powerful AI systems are advancing faster than safeguards, raising concerns that unchecked advances in AI are making it easier for armed groups to exploit the technology.