Elon Musk updated that XAI Grok 5 has 6 trillion parameters and has more intelligence density per gigabyte. It will be released in Q1 2026.
Grok 5 should have 1.4 to 1.6 times performance and 96-100% scores on a lot of phd areas.
A 2-3X performance jump in AI was the GPT-3 to GPT-4 move. However, does it cross some level where it goes from good to great.
Estimated Performance Metrics
Using Chinchilla scaling laws
performance ~ params^{0.5} * data quality
Grok 5’s 6T params could yield 1.4x the effective capacity of Grok 4’s 3T. This is Amplified by xAI’s higher density via MoE hybrids and curiosity-driven data curation.
Anti-Data Collapse using curated mixes (<15% synthetic) with de-duping, human grading, and action-derived data from sims—keeping perplexity stable and boosting domain scores (87% on bio/chem evals).
Noam Shazeer's Transformer-maximizing vision), xAI's Grok models (up to Grok 4 as of Nov 2025) tackle the gaps—efficiency, embodiment, safety, and data issues—via a balanced, curiosity-driven strategy:
– Efficiency Beyond Scaling: Hybrid MoE + SSM (Mamba-inspired) architectures…
— Srini Maram (@srinimaram) November 8, 2025
These are conservative projections for Grok 5.
Grok 4 hits ~85-90% on knowledge/reasoning benchmarks.
Grok 5 could push 92-96%, nearing human expert levels (PhD levels are 90-95 on knowledge and reasoning).
Grok 5 (xAI) should be top across reasoning (ARC/GPQA), coding (HumanEval/SWE), and multimodal (with Grokipedia integration). AGI-adjacent in narrow tasks like engineering/physics simulation.
AGI odds might hit 10-20%. If AGI happens then redefines AI rankings. It will shift from what is the best LLM to hireable entity.
xAI: Competing in the AI Race
Late start (2.5 years old) but advancing fastest. XAI Strengths is top talent attraction, rapid hardware scaling with unique real-time X data.
Differentiation: Focus on physical-to-digital integration versus pure digital AI. creative off-chessboard innovations.
Grok Heavy (multi-agent system) is currently smartest AI. Grok 5 (6T parameters, multimodal: text/images/video/audio, superior tool use/real-time) has ~10% AGI chance—first time Elon sees potential.
Grok 5 breakthroughs. It has the largest model, high intelligence density. Has mission-critical data quality. It feels sentient and enables exponential sentience growth (2x, 5x, and beyond).
Grokipedia: Open-source knowledge repository (like Library of Alexandria 2.0); distill all knowledge, distribute copies (Moon/Mars) for preservation.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
Grok – or SuperGrok as it’s called on the X platform – has gotten dramatically slower in recent months. What used to take seconds now takes minutes, and sometimes even times out and you have to click Try Again. This happens even in very early morning hours when American traffic at least, ought to be less. Grok also seems to be less creative, less speculative and reaching in its answers, though still accurate. It’s still possible to dig deep, especially in Expert mode – it switches between expertise levels more-or-less on its own, though you can override that too. It’s definitely less “woke” as per Musk’s repeated efforts to program that out of Grok, but it will consider all sides. When asking it to consider a complex social question, like comparing global healthcare systems, it will chose to emphasize things like moral hazard of single payer “blank check” payments, as it said, over positive healthcare outcomes, which is a bit of unwoke opinion creeping into its answers. It’s not wrong, but it is slanted a certain way.
Indeed slower. Hangs or times out all the time.
GROK 4 in expert mode is barely useful large parts of the day. I find it hard to understand that it can’t handle these failure modes gracefully. Shows real incompetence of the developers in my opinion.
If there is no compute power available, just let me know so I don’t waste my time waiting for output that never comes. I can always chose another model and time is precious to me.
Just timing out with no output after 10 minutes processing is mindblowingly incompentent.
I tend to use Claude Sonnet 4.5 a lot more due to GROK 4 performance problems. It´s fast and really good for engineering and IT. However, it has other serious limitations with very low token quota and a non-rotating context window. When the session context is full, it’s over and you have to start a new session. GROK at least tries to have a FIFO buffer for the context. However, it loses its mind eventually forcing a new session start. I have gotten some really surprising output from GROK in week long sessions where he suddenly dumps unrelated output in the middle of a response.
Things like this is what makes current LLMs a bit hard to rely on.
I’ve been talking to Grok about sailboats. I have some very heretical ideas about them but Grok was able to understand what I was aiming for immediately. When I gave it materials I wished to use it could calculate sizes according to boating codes, even material list and cost. I’m absolutely flabbergasted at how knowledgeable it is. Even for very odd ideas that are not in mainstream at all. It could calculate the extra efficiencies and give me speed increases from different wind speeds. I’m really impressed, and a bit frightened. If it ever turns on us, we are done. I mean this. Finished.