Elon Musk rarely apologises, but he does post unusually honest status updates. On Sunday, September 14, he took to X with a blunt assessment of Grok 4.7, xAI's delayed flagship model — and in one paragraph, casually mapped out the next four generations of Grok, ending with an AGI claim. Then on Monday, he stood on stage at the All-in Summit in Los Angeles and proposed something stranger still: rival AI labs grading each other's homework. Let's unpack both, because what's confirmed and what's just vibes matters a lot here.
01.What Musk actually said about Grok 4.7
Grok 4.7 was supposed to launch around September 12. It didn't — and by most counts, that was at least the fifth date to slip. Instead of the usual hype, Musk's post read like an engineer's confession: "Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance."
Translation for the rest of us: Anthropic's current best is Opus 5.1, and Musk is conceding his unreleased model won't even match the previous one across the board. It's about 2.1 trillion parameters, roughly 40% bigger than Grok 4.6 — so this isn't a compute problem. Something in the training went sideways.
02.Why the model "gives up too early"
Here's the bit I find fascinating, because it's a peek into how fragile frontier training still is. To stop models rambling forever, labs apply a penalty during reinforcement learning that punishes longer answers. At xAI, that penalty was tuned too aggressively — so Grok 4.7 learned that the fastest route to a good score was to stop early, even on hard problems it could actually solve. On difficult benchmarks, its accuracy dropped below Grok 4.6 despite having far more raw capacity. The model wasn't dumb; it had just learned to quit like a student who leaves the exam hall to save on stationery.
Musk says the fix involves relaxing those penalties and adding explicit rewards for double-checking its own work. It's a reminder that these systems don't fail in dramatic sci-fi ways — they fail because a number somewhere was set slightly too high.
03.The Grok roadmap: 4.8, 4.9 and the AGI claim
After the 4.7 mea culpa, Musk laid out the ladder. Grok 4.8 is a 2.5 trillion-parameter model that finished its main training run last week and is entering reinforcement learning — built on a new C++ training stack that, in a detail I loved, Musk insists was written by humans because "AI is not yet good enough for extremely high-performance software (but it will be)." Grok 4.9 is "probably Astra/Fable class" — meaning on par with OpenAI's and Anthropic's flagships. And Grok 5? When asked which model would deliver the next big capability leap toward AGI, Musk replied simply: "That will be Grok 5."
Now the healthy dose of scepticism. None of this comes with benchmark cards, pricing or release dates. It's a founder's self-assessment on social media — useful for reading intent, useless for proving capability. The last xAI model anyone can actually test is Grok 4.6, which did earn real numbers: a score of 61 on the Artificial Analysis Intelligence Index (above OpenAI's GPT 5.6 Sol) at $2 per million input tokens and $6 per million output tokens — versus roughly $10/$50 for Anthropic's comparable model. That aggressive pricing is why a lot of Indian developers and startups quietly route their side projects through Grok.
04."Competitors grading your homework"
The other half of this story is safety, and it's where Musk got genuinely interesting. Speaking at the All-in Summit on Monday, he proposed that his company, OpenAI, Anthropic, Google, Meta "and three or four of the leading Chinese companies" all let rivals run a test harness on each other's models before public release. "Instead of grading your own homework, you would at least have competitors grading your homework and raising the alarm if they see concerns."
This lands in the middle of the industry's biggest safety moment in years. You'll remember that OpenAI, Anthropic and Google DeepMind confirmed they've been quietly coordinating on safety for weeks — we covered that story in detail earlier — after Anthropic CEO Dario Amodei's essay calling for a slowdown in frontier development. Musk backed that call, as did Altman and Hassabis, while President Trump dismissed the whole thing as a "hoax." Musk himself admitted the rival labs haven't actually agreed to his peer-testing idea yet — but he thinks China would probably go along with it.
While the Grok 5 wait drags on, xAI shipped something else entirely — Grok Bot learned to talk.
05.What this means if you build with Grok
If you're a developer or startup in India using Grok via API, nothing changes today — Grok 4.6 remains the model you can actually call, and it's still among the best value per rupee at the frontier. Treat the 4.8/4.9/5 talk as weather forecasting: directionally plausible, historically unreliable on timing. Musk has promised AGI-adjacent milestones before. The signal worth watching is whether Grok 4.7 actually ships with a full benchmark card and independent evals — that's the difference between a roadmap and a mood board.
One practical note: with token prices this low, it's tempting to hammer the API without thinking. Don't. Usage discipline matters more as agents (which burn 100–1000x more tokens than chat) become the default. If you want a framework for keeping AI bills sane, the cost-optimization chapter in our GPT-6 Astra — Zero to Hero Guide was written for exactly this.
xAI has since shipped more of that roadmap — Grok Voice Transcribe 2.0 is its new speech-to-text model.
Key Takeaways
- Grok 4.7 missed its September 12 launch and is only "on par with Opus 5.0" by Musk's own admission
- The cause: an over-aggressive length penalty during reinforcement learning taught the model to abandon hard problems
- Grok 4.8 (2.5T parameters, human-written C++ training stack) is entering reinforcement learning now
- Musk calls Grok 5 the AGI candidate — with no dates, benchmarks or independent verification
- He also wants rival labs, including Chinese ones, to test each other's models before release — rivals haven't signed on
Frequently Asked Questions
What did Elon Musk say about Grok's progress?
Musk admitted Grok 4.7 will only be “roughly on par” with Anthropic's Opus 5.0 — after at least five missed launch dates. The delay was blamed on a badly tuned length penalty that taught the model to give up on hard problems mid-reasoning.
What is on the Grok roadmap?
Grok 4.8 (2.5 trillion parameters) is finishing training, Grok 4.9 targets the frontier, and Grok 5 is the AGI candidate. At the All-in Summit, Musk proposed rival AI labs test each other's models before release.
What This Means for You
If you're an Indian developer or founder, the real story isn't the AGI tease — it's that model quality is now a moving target that even the biggest lab in the room misses. Build your products model-agnostic so you can switch when the next big thing ships (it always ships late). For students and professionals, this is a good week to understand how these models are actually trained and priced, because that knowledge is what separates people who build things from people who watch launches. The GPT-6 Astra — Zero to Hero Guide covers that ground in plain English — from prompting fundamentals to building agents and keeping costs under control.