xAI's Grok 4.6/4.7 roadmap

Musk says Grok 4.6, arriving around August 7, 2026, keeps roughly the same 1.5-trillion-parameter scale as Grok 4.5 but layers in 'significantly improved' supervised fine-tuning and reinforcement learning ^[1]. Grok 4.7, due a few weeks after that, jumps to 2.1 trillion parameters and is pitched as better than 4.6 'in every way' except for a small serving-speed penalty - a tradeoff xAI is explicitly making in exchange for better token efficiency ^[1]. That efficiency framing lines up with Grok 4.5's already-public specs: roughly 80 transactions per second, $2 per million input tokens, $6 per million output tokens, and a 29.0% score on the SWE Marathon benchmark ^[2]. Read together, the roadmap suggests xAI is now optimizing as much for tokens-per-dollar and serving throughput as for raw parameter count.

Every spec in this story - both parameter counts, the SFT/RL claims, and the serving-speed-for-efficiency tradeoff - traces back to a single Elon Musk post on X; xAI itself has published no launch page, pricing, or benchmark data for either model ^[3]. The numbers aren't even internally consistent across the outlets covering them: one aggregator's writeup describes Grok 4.6 in one line as a '150-billion-parameter' model while repeating the '1.5-trillion-parameter' figure elsewhere in the same piece, which reads as a transcription error rather than a genuine alternate spec, but it underscores how thin the verification layer around this announcement is ^[4]. Even Musk's own timeline was casual rather than a firm date - 'Grok 4.6 in 2 weeks and Grok 4.7 in 4 weeks' - and it's third-party outlets, not xAI, doing the math to turn that into calendar dates ^[3].

On X, the announcement is being rebroadcast almost verbatim by fast-turnaround accounts with little independent analysis or pushback - no skepticism surfaced in the sampled reaction there. Reddit's read is far more guarded: commenters widely treat Musk's release-date promises as unreliable, pointing to his track record of missed timelines in other ventures, and several dispute the parameter counts themselves - some believe Grok 4.6 is closer to 2 trillion parameters, which would directly contradict the 1.5-trillion figure in Musk's post. There's also a live side debate over whether Grok 4.5, the base these new models are meant to improve on, is an original xAI architecture or a fine-tuned derivative of another lab's model - unresolved, and worth treating as community speculation rather than fact.

Coverage frames the back-to-back 4.6/4.7 releases as an unusually compressed cadence - three substantively different frontier models (4.5, 4.6, 4.7) shipping within about two months, positioned against OpenAI, Google, and Anthropic's typically slower release rhythms ^[5]. It's also a notably informal way to announce a frontier model: the roadmap arrived as a short reply on X rather than a blog post or launch event, just days after Grok 4.5 itself landed in mid-July 2026 ^[6]. Whether that speed reflects genuine engineering velocity or front-loaded marketing is exactly the question the model's eventual (still absent) benchmarks will have to answer.