Musk shelves Grok 4.7, unveils 2.5T-parameter Grok 4.8, teases Grok 5 as first AGI

Elon Musk laid out an aggressive Grok roadmap, shelving the never-publicly-released Grok 4.7 and unveiling Grok 4.8 — a 2.5-trillion-parameter model trained on xAI's new C++ software stack that Musk said will finish pretraining this week and then enter reinforcement learning. He characterized Grok 4.7 as roughly equivalent to Opus 5.0, placed Grok 4.9 in the Astra/Fable capability class, and teased Grok 5 as 'maybe better than anything' — explicitly framing it as xAI's first AGI-capable model, though he offered no benchmarks or timeline. The current publicly listed flagship remains Grok 4.6 with a 500,000-token context window.

The version churn drew immediate skepticism. Musk's unbenchmarked AGI claim landed awkwardly against his simultaneous public support for Dario Amodei's 'Pace the Frontier' slowdown proposal — critics on X noted the tension between backing an industry slowdown and racing to ship a trillion-parameter model and declare imminent AGI. The shifting version numbers (shelving 4.7 while announcing 4.8) added to the sense of roadmap-by-tweet.

Separately, Grok models were added to Microsoft's Copilot Frontier program in M365, with Satya Nadella posting 'More model choice coming to Copilot. Welcome Grok!' — a notable distribution win that puts xAI's models alongside OpenAI's and Anthropic's inside Microsoft's enterprise suite.

What to watch: whether Grok 4.8 ships with real benchmarks after RL, whether the C++ stack delivers measurable training-efficiency gains, and whether the AGI framing is substantive or marketing ahead of xAI's next fundraising or compute build-out.