Grok 5: Release Date, Specs, and Elon Musk's AGI Claims Explained
Elon Musk says Grok 5 is a 6 trillion parameter model with a shot at AGI. Here's what's confirmed about specs, timeline, and benchmarks.
What is Grok 5?
Grok 5 is the next frontier AI model from Elon Musk’s xAI, currently in training on the company’s Colossus 2 supercomputer in Memphis. Musk has described it as a 6 trillion parameter model, twice the size of the 3 trillion parameter base behind Grok 3 and Grok 4, built to be multimodal from the ground up (text, images, video, and audio) and tuned for real-time video understanding and tool use. Musk has also said it’s the first Grok model he thinks has a real, if small, chance at artificial general intelligence.
TL;DR
- Grok 5 is reportedly a 6 trillion parameter model, double the 3 trillion parameter base used for Grok 3 and Grok 4, according to Musk.
- Musk has put a specific number on AGI odds, saying he calculates roughly a 10% chance Grok 5 reaches artificial general intelligence, with a 90% chance it doesn’t.
- The model is being trained on Colossus 2, xAI’s second Memphis supercomputer, and xAI has said the next major Grok model will draw on SpaceX’s own launch, engineering, and manufacturing data.
- Musk has emphasized real-time video understanding as a capability he considers essential for AGI, something no current model does well.
- Musk’s own benchmark prediction is that Grok 5 could score “nearly perfect” on Humanity’s Last Exam, one of the hardest text benchmarks available today.
- The original target was Q1 2025, but the timeline given on SpaceX’s earnings call has since shifted to before the end of 2026.
- xAI currently trails the frontier on independent rankings, so Grok 5 has a real gap to close before any of these claims can be verified.
What has Elon Musk actually said about Grok 5’s specs?
Musk has made a handful of specific claims about Grok 5’s architecture, though none of them come with official documentation yet. He says Grok 5 is built on 6 trillion parameters, compared to the 3 trillion parameter base used for Grok 3 and Grok 4. Parameters are the internal values a model adjusts during training, and more of them generally means more capacity to store knowledge and represent complex patterns, though size alone doesn’t guarantee better performance.
Musk has also talked about “intelligence density per gigabyte,” a framing that matters because memory and compute are expensive. A model that packs more capability into less memory is cheaper and faster to run in production. He’s claimed Grok 5 will have much higher intelligence density than Grok 4, partly thanks to better training data. On SpaceX’s first earnings call, Musk said the next major Grok model would train on SpaceX’s own internal data, including launch logs, engineering records, and manufacturing data, a dataset no other AI lab has access to.
Grok 5 is also described as inherently multimodal, meaning it’s designed from the start to handle text, images, video, and audio together, rather than bolting vision or audio onto a text-first model later.
Why does Musk think Grok 5 has a shot at AGI?
Musk has said Grok 5 is the first model where he thinks xAI has a non-zero chance of reaching artificial general intelligence, the idea of an AI that can perform essentially any cognitive task a human can. He put a number on it: roughly 10%, which he was careful to frame as still meaning a 90% chance it doesn’t happen.
The capability Musk keeps returning to is real-time video understanding. Most current AI models can analyze still images or short pre-recorded clips, but processing a live video feed as it happens is a much harder technical problem. Musk has argued that without this ability, a model can’t achieve AGI, since humans process the visual world in real time as a matter of course. He’s also pointed to improved tool use, including the ability to create new tools on the fly, as a core part of what separates Grok 5 from earlier models. Musk added that xAI has unspecified “special sauce” features in the works that he says will make the model “really feel sentient,” though he didn’t go into detail on what those are.
What benchmark scores does Grok 5 need to hit?
The current frontier, measured by the Artificial Analysis Intelligence Index (an independent composite score covering reasoning, coding, science, and math), is tight. Gemini 4 Carbon currently sits at the top with a score of 59, with Anthropic’s and OpenAI’s best models close behind. Grok’s most recent model, Grok 4.7, scored 46 on the prior version of that index, well behind the pack. On Terminal-Bench, a test of AI agents performing real coding tasks on a computer, Grok 4.7 scored around 26%, compared to roughly 60% for GPT-6 Astra.
That’s the gap Grok 5 needs to close. Musk’s own prediction is far more aggressive than catching up: he’s said Grok 5 might score “nearly perfect” on Humanity’s Last Exam, a benchmark made up of extremely difficult, expert-written questions designed to resist easy AI solutions. For comparison, Musk has said Grok 4, which he called “primitive” by comparison, scored around 52% on the text portion of that same exam. He’s also said Grok 4.9, the model expected to arrive before Grok 5, should already be competitive with GPT-6 Astra and Claude-class models, which would make Grok 5 the one expected to clearly surpass them.
A realistic bar for Grok 5 to be taken seriously: matching or beating Gemini 4 Carbon’s roughly 60 on the Artificial Analysis Index, closing the agentic coding gap seen on Terminal-Bench, and standing out specifically on video and multimodal tasks, where Musk claims Grok 5 will have an edge nobody else has.
When is Grok 5 actually coming out?
The timeline has already moved once. Musk’s original target, stated when training began, was the first quarter of 2025. That date has since slipped. On SpaceX’s earnings call, the company gave a new target of before the end of 2026.
Part of the delay comes from how model training works. After the initial large-scale training run, models go through reinforcement learning, where they practice tasks and get rewarded for good performance. This stage often takes longer than expected and is where a lot of real-world polish happens. xAI has also said it rebuilt its training system in C++, a project largely handled by engineers from SpaceX’s Starlink software team, which can improve future models but adds time upfront. Grok 4.8, an earlier model, has already missed its own stated release window, which gives some sense of how fluid these timelines tend to be in practice.
Is xAI actually positioned to compete at the frontier?
It’s a fair question, given how far behind Grok currently sits on independent benchmarks. But the compute buildout behind Grok 5 is real and large. Colossus, xAI’s first Memphis supercomputer, was reportedly built in 122 days. On SpaceX’s earnings call, the company said it had around 1.4 gigawatts of computing power online, with a target of 2 gigawatts by the end of the year and 10 to 15 gigawatts by the end of 2027. For scale, 1 gigawatt is roughly the output of a large power station. SpaceX reported close to 20 billion dollars in capital spending in a single recent quarter, most of it going toward AI infrastructure, and the company has had enough spare compute to rent some out to other AI firms, including Anthropic.
Musk’s argument, made directly to podcaster Dwarkesh Patel, is that algorithmic ideas spread between AI labs within about six months regardless of who discovers them first, so the long-term advantage goes to whoever can scale hardware fastest. That’s a bet that xAI’s strength in building infrastructure (inherited partly from SpaceX and Tesla’s manufacturing culture) matters more than any single research breakthrough. It’s also worth remembering how quickly rankings can shift: Google went from being widely written off to taking the top spot on the Artificial Analysis Index within about a month, releasing two major models in quick succession. That kind of swing shows the leaderboard isn’t fixed, which cuts both ways for xAI: it means Grok 5 could leapfrog the field, but also that whatever it lands on may not hold the top spot for long.
Frequently Asked Questions
How many parameters does Grok 5 have?
Musk has said Grok 5 is a 6 trillion parameter model, double the 3 trillion parameter base used for Grok 3 and Grok 4. This figure comes from Musk’s own public statements, not an official technical paper, so it should be treated as unverified until xAI publishes documentation.
When will Grok 5 be released?
Musk originally targeted the first quarter of 2025 when training began. The timeline has since shifted, with SpaceX’s earnings call setting a new target of before the end of 2026.
Does Grok 5 really have a chance at AGI?
Musk has said he estimates roughly a 10% chance Grok 5 reaches artificial general intelligence, which also means a 90% chance it doesn’t. It’s the first Grok model he’s described in these terms, but it remains a personal estimate from Musk, not a measured outcome.
What makes Grok 5 different from Grok 4?
Besides being roughly twice the parameter count, Grok 5 is designed to be multimodal from the start (text, image, video, audio), trained on higher-quality data including SpaceX’s internal engineering and launch data, and built with an emphasis on real-time video understanding and tool use that Musk says current models lack.
How does Grok 5 compare to GPT-6 Astra and Gemini 4 Carbon?
Based on current benchmarks, xAI’s latest released model, Grok 4.7, trails both, scoring 46 on the Artificial Analysis Intelligence Index compared to 59 for Gemini 4 Carbon, and far behind on agentic coding tests like Terminal-Bench. Grok 5 has not been released or independently benchmarked yet.
