xAI is shipping Grok 4.6 and 4.7 back to back in a release cadence no frontier lab has matched

XAI confirmed a 2-trillion-parameter Grok 4.6 on July 18, 2026, with a larger Grok 4.7 already weeks behind it, as Elon Musk openly targets Moonshot AI's Kimi K3 that wiped $3.3 trillion from chip stocks days earlier. The release cadence of three frontier models in under two months, powered by xAI's Colossus supercluster with 555,000 GPUs, is unmatched by any other lab. Grok 4.5 launched on July 8 at $2 per million input tokens, over 60 percent cheaper than Anthropic's Opus 4.8 and GPT-5.5, while scoring highest on agentic tool use.

xAI confirmed a 2-trillion-parameter Grok 4.6 on July 18, 2026, with a larger successor already weeks behind it, as Elon Musk openly targets Moonshot AI's Kimi K3 that wiped $3.3 trillion from chip stocks days earlier. Elon Musk announced on July 18 that Grok 4.6, a 2-trillion-parameter model, had completed initial pre-training the week of July 20 and could "exceed Kimi" while running faster and more token-efficiently. That is a direct call-out of Moonshot AI's Kimi K3, the 2.8-trillion-parameter open-weight model that triggered semiconductor selloffs comparable to the DeepSeek shock of early 2025. The Philadelphia Semiconductor Index fell nearly 10 percent in a single week. Moonshot's own servers buckled under demand so quickly the company temporarily stopped accepting new subscribers. And xAI's response is: here comes something bigger. What makes this remarkable isn't the model count. It's the pace. Grok 4.5, a 1.5-trillion-parameter model, launched on July 8 at $2 per million input tokens and $6 per million output tokens, with a 500,000-token context window and built-in web search. That's over 60 percent cheaper than Anthropic's Opus 4.8 and GPT-5.5. Grok 4.6 at 2 trillion parameters is confirmed and in post-training now. A follow-on Grok 4.7, estimated in the range of 4 to 6 trillion parameters, is already on the roadmap weeks behind it. Three substantively different frontier models in under two months. No other lab has done this publicly. The infrastructure behind this cadence is the real story. xAI's Colossus supercluster in Memphis had expanded to 555,000 GPUs as of January 2026, integrating Nvidia H200s alongside the newest Blackwell GB200 and GB300 units, with a projected buildout toward 1 million GPUs and 2 gigawatts of power. That is roughly 100,000 more GPUs than Microsoft's first Blackwell campus in Abilene, Texas, which came online in early 2026 as part of Project Stargate. OpenAI relies heavily on that Microsoft infrastructure. Anthropic, interestingly, signed a deal in May 2026 for exclusive access to Colossus 1, the older 220,000-GPU Memphis cluster, which xAI had already outgrown. xAI is training on what comes after that. The implication is straightforward: when you own the compute at this scale and you've already moved on from the cluster you leased to a competitor, you can run training runs in parallel or in rapid sequence that others simply can't. Monthly foundation model releases, which xAI has stated as a target through the end of 2026, stop being a marketing claim and start being plausible engineering when you have a million GPUs pointed at it. For enterprise developers, the pricing dynamic is the pressure point. Grok 4.5 at $2 per million tokens with cached input at $0.50 lands significantly below what OpenAI and Anthropic charge for comparable capability tiers. On Artificial Analysis's Intelligence Index, Grok 4.5 scores 54 and ranks fourth overall, behind Claude Fable 5, Opus 4.8, and GPT-5.5, but it tops the field on agentic tool use. That last point matters more than the aggregate score for the developers building autonomous agents, the fastest-growing segment of API consumption. If Grok 4.6 improves on that ranking while holding the pricing floor, the renewal-cycle math changes for any team that hasn't yet locked into a long-term OpenAI or Anthropic contract. The Kimi K3 target is telling Musk didn't mention OpenAI or Anthropic when he framed 4.6's ambition. He mentioned Kimi K3. That's a deliberate choice, and it points at where xAI sees the real competitive threat right now. Kimi K3 is open-weight, meaning developers can run it themselves without paying Moonshot's API. According to Bloomberg, its advantage may lie more in memory architecture than raw compute, which is what makes it efficient enough to rattle chip stocks even at 2.8 trillion parameters. If xAI ships a 2-trillion-parameter model that matches or beats Kimi K3's performance at a fraction of the cost, while remaining on a closed, enterprise-grade API with guaranteed uptime, it closes the main practical argument for choosing the open-weight alternative. Wall Street's reassessment of the Kimi shock is worth noting here. After the initial $3.3 trillion selloff, Bank of America, UBS, and Morgan Stanley each argued that low-cost, high-capability models drive more inference volume, not less, which means more chip demand over time. That logic benefits xAI too, since more inference demand at aggressive prices still runs on Colossus. Moonshot, for its part, responded to the moment by circulating a shareholder resolution seeking approval for a Hong Kong IPO at a valuation above $30 billion. That's the market Musk is moving into. Grok 4.6 hasn't launched publicly yet. No independent benchmarks exist, no context window is confirmed, and Musk's projections about beating Kimi are, for now, projections. But the cadence itself is already doing work. Every week that passes between Grok 4.5's launch and 4.6's release is a week OpenAI and Anthropic spend defending renewal conversations with enterprise clients who now have a cheaper, faster-releasing alternative asking for the deal. That pressure doesn't require 4.6 to win every benchmark. It only requires it to show up. Also read: The hedge funds that made 60% on the AI chip rally are now down 17% in July as $137 billion flees Asia https://startupfortune.com/the-hedge-funds-that-made-60-on-the-ai-chip-rally-are-now-down-17-in-july-as-137-billion-flees-asia/ • The EPA wants to strip neighbors of their right to object to data center pollution permits https://startupfortune.com/the-epa-wants-to-strip-neighbors-of-their-right-to-object-to-data-center-pollution-permits/ • Nvidia is putting $1 billion into Naver as its Korea sovereign AI push shifts from handshakes to hard cash https://startupfortune.com/nvidia-is-putting-1-billion-into-naver-as-its-korea-sovereign-ai-push-shifts-from-handshakes-to-hard-cash/