Quick Brief
- Grok 5 is actively training on Colossus 2, the world’s first gigawatt-scale AI cluster, confirmed operational in January 2026
- The model carries 6 trillion parameters in a Mixture-of-Experts architecture, making it the largest publicly announced AI model
- Elon Musk assigned a 10% and rising probability that Grok 5 reaches human-level AGI performance
- A public beta is estimated for March to April 2026, with full API access expected in Q2 2026
Grok 5 remains in active training, and every additional week on Colossus 2 is telling you something specific: xAI is not rushing this one. What you need to understand is not just when it arrives, but what the extended compute window actually signals about what this model will be capable of doing.
Why Colossus 2 Changes the Entire Scale of This Release
xAI officially activated Colossus 2 in January 2026, confirming it as the first gigawatt-scale AI training cluster in the world. Elon Musk announced the milestone directly on X, confirming the facility had crossed the 1 GW threshold and is targeting an upgrade to 1.5 GW by April 2026. The Memphis site spans three buildings, with the third structure purchased in December 2025 and targeted for conversion in Q1 2026.
The total Colossus complex targets 555,000 NVIDIA GPUs, representing roughly an $18 billion GPU investment at approximately $32,400 per unit. Colossus 1 alone houses 230,000 GPUs including 32,000 GB200s; Colossus 2 adds 550,000 GB200s and GB300s as they come online. NVIDIA CEO Jensen Huang described the original Colossus buildout as “superhuman,” noting it went from construction to operational in 19 days, compared to a traditional 4-year timeline. The 2 GW total capacity makes the site 4x more powerful than the next-largest dedicated AI training facility globally.
What 6 Trillion Parameters Actually Means
Grok 5 carries 6 trillion parameters in a Mixture-of-Experts (MoE) architecture, making it the largest publicly announced AI model by parameter count. In an MoE design, only a fraction of those parameters activate per query, which means the model gains massive capacity without proportionally massive inference costs. Musk confirmed the 6 trillion figure at the Baron Capital conference in November 2025, describing the model as having higher “intelligence density per gigabyte” compared to its predecessor.
For context, Grok 4 carried an estimated 3 trillion parameters. Grok 5 doubles that scale while introducing native multimodal architecture from the ground up, covering text, images, audio, and real-time video with temporal reasoning. The reported context window extends to 1.5 million tokens, a significant jump from Grok 4’s 128K standard and 2M extended configurations.
The Release Window: What the Evidence Points To
Musk confirmed in November 2025 that Grok 5 was delayed from late 2025 to Q1 2026, citing the decision to scale the model significantly before release. That Q1 2026 window was not met. As of February 25, 2026, Grok’s own X account updated the projection, pointing to Q2 2026 as the most likely window for full release.
The most credible independent analysis, based on xAI’s rollout patterns, places a public beta between March and April 2026, followed by full API access in Q2 2026. This timeline also aligns with the Colossus 2 upgrade to 1.5 GW confirmed for April 2026, which may mark the completion of the primary training run.
Why the H1 vs H2 Distinction Matters
A H1 2026 release keeps Grok 5 competitive against deployed frontier models that are already accumulating developer integrations and enterprise workflows. Grok 4.20 Beta launched February 17, 2026 with a 4-agent collaboration system (Grok, Harper, Benjamin, Lucas), followed by Grok 4.20 Heavy on February 18 with 16 specialized agents. These releases appear to serve as a live test of the multi-agent architecture that Grok 5 is designed to run natively and at scale.
Grok 5 Against the Current Frontier
Grok 5 is not yet released, but the existing competitive context defines the performance bar it is being built to clear.
| Model | Parameters | Context Window | Architecture | Confirmed Benchmark |
|---|---|---|---|---|
| Grok 4 | ~3T (reported) | 128K standard / 2M extended | MoE | ARC-AGI 92.7%, Chatbot Arena |
| Grok 4.1 | Not disclosed | Standard | MoE optimized | API: $0.20/$0.50 per 1M tokens |
| Claude Opus 4.6 | Not disclosed | 1M tokens | Dense transformer | SWE-Bench 77.2% |
| Gemini 3.1 Pro | Not disclosed | 1M tokens | MoE | ARC-AGI-2 77.1% |
| Grok 5 (in training) | 6T confirmed | 1.5M tokens (reported) | MoE | Not yet published |
Grok 4’s ARC-AGI score of 92.7% on Chatbot Arena already placed it among top-tier reasoning models before Grok 5 enters the picture. Grok 4.1 Fast also currently holds the lowest price point among frontier-class models at $0.20 per million input tokens, a competitive dynamic that benefits developers regardless of which model they choose.
The Multi-Agent Architecture: From 4.20 to 5
The Grok 4.20 releases in February 2026 provide the clearest structural preview of what Grok 5 will do at a larger scale. Grok 4.20 Beta introduced 4-agent collaboration for research and analysis tasks; Grok 4.20 Heavy scaled that to 16 specialized agents for deep research at the SuperGrok Heavy tier ($300/month). Grok 5 is designed to extend this further through dynamic agent spawning, where the number of agents scales based on task complexity rather than being fixed.
Persistent memory across agent sessions and real-time data access through the X live stream are also confirmed targets. This live X integration is architecturally unique: no other frontier model has native access to a real-time social data stream as part of its core inference pipeline.
The AGI Question: What a 10% Probability Actually Signals
Elon Musk stated at the Baron Capital conference in October to November 2025 that Grok 5 carries a 10% and rising probability of reaching human-level AGI performance. This is not a standard marketing claim. No prior Grok generation carried a comparable internal estimate, and the capability targets cited alongside it are specific: real-time video reasoning, multi-step autonomous task execution, and performance exceeding human experts in engineering and research domains.
Limitations and Honest Considerations
Musk’s product timelines have shifted multiple times. Grok 5 was originally targeted for late 2025, then Q1 2026, and neither window was met as of February 2026. The 10% AGI probability lacks a verifiable benchmark definition, and other major labs including Anthropic and DeepMind, working at comparable scale, have not made similar claims. Parameter count alone has historically not produced qualitative intelligence leaps in prior generations. The more meaningful signal may be the rapid-learning architecture introduced in Grok 4.20, which enables weekly model updates using real-world feedback rather than requiring full retraining.
What to Watch Before the Release
Three signals will mark meaningful progress toward launch:
- Grok 4.20 exiting beta with official benchmarks, targeted for March 2026, which will provide a baseline for evaluating the step up to Grok 5
- Colossus 2 power upgrade to 1.5 GW confirmed for April 2026, which may coincide with the final training phase
- Any official benchmark disclosure from xAI against Gemini 3.1 Pro or Claude Opus 4.6, signaling the model has reached a publishable performance level
Frequently Asked Questions (FAQs)
When will Grok 5 be released?
As of February 26, 2026, Grok 5 has no confirmed public release date. The Q1 2026 window was not met. Grok’s own X account updated the projection on February 25, 2026 to Q2 2026 for full release, with a public beta estimated between March and April 2026.
What is Colossus 2 and why does it matter for Grok 5?
Colossus 2 is xAI’s gigawatt-scale AI supercomputer in Memphis, Tennessee, confirmed operational in January 2026 as the first training cluster in the world to surpass 1 GW. It is being upgraded to 1.5 GW by April 2026 and houses the infrastructure on which Grok 5 is being trained.
How many parameters does Grok 5 have?
Grok 5 carries 6 trillion parameters in a Mixture-of-Experts architecture. Elon Musk confirmed this figure at the Baron Capital conference in November 2025. The MoE design means only a subset of parameters activates per query, keeping inference costs manageable despite the massive scale.
How many GPUs does the Colossus complex hold?
The total Colossus complex targets 555,000 NVIDIA GPUs across three Memphis buildings. Colossus 1 is operational with 230,000 GPUs including 32,000 GB200s. Colossus 2 and the third building add 550,000 GB200s and GB300s, representing roughly $18 billion in GPU investment.
Can Grok 5 actually achieve AGI?
Elon Musk stated a 10% and rising probability at the Baron Capital conference in late 2025. The claim is not backed by a universally accepted AGI benchmark, and other leading labs at comparable scale have not made similar public claims. It reflects ambition more than a verified technical threshold.
Why was the Grok 5 launch delayed from 2025?
Musk confirmed in November 2025 that the delay was a deliberate decision to make the model significantly larger before release. Scaling to 6 trillion parameters and building the Colossus 2 infrastructure to support that training run required more time than the original late-2025 schedule allowed.
What multimodal capabilities will Grok 5 support?
Grok 5 is designed as a natively multimodal model from the ground up, covering text, images, audio, and real-time video with temporal reasoning. Unlike earlier Grok versions where vision was added post-architecture, multimodal processing is built into the core model design.
What platforms and pricing will Grok 5 use?
Based on xAI’s established pattern, Grok 5 is expected to launch first for X Premium+ and SuperGrok subscribers. A tiered API pricing structure is anticipated, consistent with Grok 4’s $3 input/$15 output per million tokens, with a lower-cost Fast variant likely to follow.
