xAI SKIPPED Grok 4.7… Here’s What’s Coming Instead

Link to our newsletter: https://bitbiased.ai/ Grok 4.7 was built to persist through difficult, long-running tasks. According to Elon Musk, its reinforcement learning may have accidentally taught it to do the opposite: give up early. Then, before Grok 4.7 even shipped, Musk started talking about Grok 4.8 — a reportedly 2.5-trillion-parameter model trained on a brand-new C++ software stack. That raises a bigger question: is xAI genuinely accelerating toward the frontier, or are the model numbers moving faster than the evidence? Grok 4.6 gives us the clearest baseline. Released August 12, 2026, it brought a 500,000-token context window, configurable reasoning, text and image input, function calling, structured output, and aggressive API pricing at $2 per million input tokens and $6 per million output tokens. xAI also reported meaningful gains across agentic benchmarks. Grok 4.6 reached 69.9% on CursorBench, 65.9% on DeepSWE, 57.5% on APEX-Agents, and 26% on Terminal-Bench 3.0. That last result is especially important because xAI's roadmap has increasingly focused on models that can keep working through complex engineering and professional tasks rather than simply producing better chatbot answers. Grok 4.7 was supposed to push that further. Musk described it as roughly 2.1 trillion parameters, compared with approximately 1.5 trillion for 4.6. But its release window slipped repeatedly. On September 11, Musk finally gave a technical explanation: reinforcement learning may have penalized long answers too aggressively, causing the model to abandon difficult problems early and insufficiently check its own work. That's almost the exact opposite of what a persistent AI agent needs. And that's where Grok 4.8 enters the story. On September 13, Musk described Grok 4.8 as a 2.5-trillion-parameter model running on a new low-level software stack. But there is an important catch: parameter counts aren't necessarily directly comparable. Grok-1, for example, used a mixture-of-experts architecture where only part of its total parameter count was active for each token. Without architectural details for Grok 4.8, 2.5 trillion tells us much less about real inference compute or capability than it initially appears to. The potentially bigger development is xAI's software stack. Musk has previously discussed an in-house low-level training stack designed for enormous GPU clusters and claimed major performance advantages over JAX on very large runs. The engineering idea is plausible; the claimed magnitude has not been independently demonstrated. That matters because xAI's compute story isn't simply about owning GPUs. Colossus has expanded dramatically, but efficiently coordinating huge numbers of accelerators — especially across infrastructure and networking constraints — is its own engineering problem. A better training stack could potentially matter more than adding another trillion parameters. Meanwhile, the competitive bar keeps moving. GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash are all part of the landscape Grok's next models have to compete against. This video separates what xAI has actually shipped from what Musk has announced, explains what went wrong with Grok 4.7, breaks down what the 2.5-trillion-parameter Grok 4.8 claim really tells us, and looks at why xAI's new C++ stack may ultimately be the more important story. CHAPTERS 00:00 Grok 4.7 Learned to Give Up 01:54 What Grok 4.6 Actually Is 05:07 Where xAI Was Already Headed 06:00 The Grok 4.7 Promise 06:46 What Actually Went Wrong With 4.7 08:56 Then Grok 4.8 Showed Up — Before 4.7 Even Shipped 10:31 Why the C++ Stack Might Be the Real Story 11:47 The Compute Behind the Claim — And Its Limits 14:07 What Grok 4.8 Actually Has to Beat 15:22 The Actual Verdict #grok #grok48 #xai #elonmusk #ai