Musk Confirms Colossus 2 Trained New Grok Update, Teases Agent Mode Unlock

Elon Musk announced a new Grok release today, May 18, 2026, confirming for the first time that xAI's Colossus 2 supercomputer cluster was used to partially train the model. Musk posted directly on X inviting users to try the update and flagged a new Agent Mode as what he called a "major ability unlock."

The release is xAI's most consequential model push since Grok 4.20 in early May, and it lands one day before Google's Gemini 4.0 reveal at I/O 2026 — timing that is almost certainly deliberate.

What Changed in the New Grok

The update is positioned as an incremental release rather than the long-promised Grok 5, but the technical claims are notable. Musk says Agent Mode lets Grok handle significantly more complex, multi-step autonomous tasks — moving beyond simple Q&A into proactive, goal-driven execution. Practically, that means longer-running tool calls, multi-app workflows, and persistent state across sessions.

The new model inherits the 2-million-token context window that Grok 4.20 introduced on May 1 — one of the largest available from any AI provider — and adds native multimodal support across text, images, audio, and video.

Why this matters: For xAI, Agent Mode is the company's first real entry into the agentic coding and workflow market dominated by Claude Opus 4.7 and OpenAI's Codex. The 2-million-token window is competitive on paper with Google's expected Gemini 4.0 announcement tomorrow, and gives xAI a credible pitch for long-context document analysis.

Colossus 2 Comes Online

Colossus 2 is xAI's second-generation training cluster, built to address shortcomings of the original Colossus supercomputer. Tom's Hardware reported earlier this year that Colossus 1's mixed-GPU architecture proved inefficient for frontier training runs — to the point where xAI rented Colossus 1 capacity to Anthropic for inference workloads while building a unified Blackwell-only Colossus 2 for frontier training.

xAI has claimed Colossus 2 will eventually reach 1 gigawatt of capacity with 1 million GPUs. Satellite imagery analyzed by Tom's Hardware suggests the actual site capacity sits closer to 350 megawatts of cooling — meaning xAI is either staging the buildout or the public claims are aspirational. Either way, today's announcement is the first confirmed Grok model trained even partially on the new cluster.

For context: Anthropic disclosed a 3.5-gigawatt TPU commitment with Google earlier this year, and OpenAI's combined cluster footprint across Microsoft, Oracle, and SoftBank is reportedly in a similar range. Compute is now the gating factor on frontier models, and xAI has been catching up rather than leading.

Where Agent Mode Sits Against Rivals

Agent Mode arrives in a crowded market. Anthropic's Claude Agent SDK lets developers build long-running agents with tool use and persistent memory. OpenAI's ChatGPT Codex now ships native agentic coding on iOS, Android, and the web. Google is expected to make agentic coding a centerpiece of tomorrow's I/O developer keynote.

Where xAI has a distribution advantage is X integration. Grok with Agent Mode can act on a user's X account, draft posts, monitor mentions, and trigger workflows tied to the platform — capabilities none of the major competitors can match without API access. For X power users and political accounts, that integration is the real product differentiation.

Expert Perspectives

Reactions on X this morning have been split. Bulls point to the speed at which xAI has shipped capability — Grok went from a generic chatbot in 2023 to a frontier-competitive model with agentic features in roughly 30 months. Skeptics note that public benchmarks for Grok have consistently trailed Anthropic and OpenAI by 3 to 6 months, and that xAI's marketing has outrun the published evaluations.

The other tension: xAI is preparing for an IPO, with Wall Street pitches reportedly handled by Apollo and Morgan Stanley in coordination with the SpaceX IPO machinery. A high-profile Colossus 2 announcement timed against Google I/O is exactly the kind of narrative push a pre-IPO company runs.

What's Next

Three things to watch over the next two weeks. First, public benchmark numbers for the new Grok — if SWE-bench Pro and Terminal-Bench scores ship alongside the model, xAI is confident; if they do not, the marketing has outrun the substance. Second, Colossus 2's actual power draw and GPU count, which satellite analysts will publish over the summer. Third, the Grok 5 release window, which Musk previously suggested would be a public beta in May or June 2026 with full API access in Q3.

For now, the new model is rolling out across X Premium tiers and the standalone Grok app on iOS and Android.

The Bottom Line

Today's Grok update is the first sign that xAI's compute buildout is producing usable frontier models rather than just press releases. Agent Mode brings xAI into the same product category as Claude and ChatGPT, and Colossus 2 — even at 350 megawatts — is real infrastructure. Whether Grok closes the benchmark gap with Anthropic and OpenAI is still an open question, but the company is no longer fighting from behind on infrastructure.