Elon Musk says Grok 5 will be a 6 trillion parameter model, but by early August 2026 it had not shipped. For buyers, size stopped predicting usefulness: active parameters, cost per finished task and accuracy on your own work decide results.
Not much, and less each year. Elon Musk has said Grok 5 will be a 6 trillion parameter model, but by early August 2026 it had not been released, and the newest Grok on sale was Grok 4.5, launched on 8 July. For a business buyer, results depend on how much of a model works on each request, what a finished task costs and how often the answer is right, not on headline size.
Six trillion is a number built for repetition. But many large models now use a mixture-of-experts design that switches on only a slice of their weights for each word, the gap between leading models has narrowed, and the price of good-enough intelligence has collapsed. Here is how to read size claims, and what to test instead.
- parameters Musk said Grok 5 would have, in a November 2025 interview
- 6T
- parameters Musk said Grok 5 would have, in a November 2025 interview
- of Kimi K2's 1 trillion parameters are active for each token
- 3.2%
- of Kimi K2's 1 trillion parameters are active for each token
- fall in the cost of GPT-3.5-level AI, November 2022 to October 2024
- 280x
- fall in the cost of GPT-3.5-level AI, November 2022 to October 2024
- score gap between the top two models in Stanford's 2025 AI Index
- 0.7%
- score gap between the top two models in Stanford's 2025 AI Index
01
What was promised, and what shipped
The 6 trillion figure is a statement, not a specification sheet. In a November 2025 interview with the investor Ron Baron, Musk described Grok 5 as a 6 trillion parameter model and said Grok 3 and Grok 4 were based on a 3 trillion parameter model. He did not say how many of those parameters would work on each request, and that is the figure that matters for speed and cost, so it is the one to ask for.
Nov 2025
Grok 5 described at 6 trillion parameters
Musk tells Ron Baron that Grok 5 doubles the 3 trillion parameter base behind Grok 3 and 4.
8 Jul 2026
Grok 4.5 ships
SpaceXAI's first model since the company went public, at $2 / $6 per million tokens. Musk calls it Opus-class.
28 Jul 2026
A new roadmap
Musk posts that Grok 4.6, a 1.5 trillion parameter model, will release around 7 August, with a 2.1 trillion parameter Grok 4.7 a few weeks later.
5 Aug 2026
Grok 5 still unreleased
The next Groks on Musk's own schedule are 4.6 and 4.7.
02
Total parameters versus the ones that do the work
Many very large models now use a mixture-of-experts design. The model is split into many specialist sub-networks, and a router sends each token, roughly a word or part of one, to a few of them. Total parameters describe capacity. Active parameters describe the work done, and the cost, for each token.
Kimi K2, an open model from Moonshot AI, shows the gap. It has 1 trillion parameters and 384 experts, and it routes each token to 8 of those experts plus one shared expert that is always on, so only 32 billion parameters are active at a time. The published model cards below tell the same story.
| Model | Total parameters | Active per token |
|---|---|---|
| Kimi K2 (Moonshot AI) | 1 trillion | 32 billion |
| DeepSeek-V3 | 671 billion | 37 billion |
| Qwen3-235B-A22B (Alibaba) | 235 billion | 22 billion |
| gpt-oss-120b (OpenAI) | 117 billion | 5.1 billion |
| Llama 4 Scout (Meta) | 109 billion | 17 billion |
So a bigger total does not automatically mean a slower or dearer model: the work done for each token follows the active count. All the parameters still have to fit in memory, though, which is why gpt-oss-120b needs a single 80GB GPU even with only 5.1 billion active. Nor does a big total guarantee quality. Training data, routing and tuning decide that, which is why two models of similar size can behave very differently on the same job.
Most closed labs publish even less. Anthropic's model overview lists prices, context windows and knowledge cut-offs for its current Claude models, but no parameter counts. When most of the companies that build these models do not lead with size, buyers need not either.
03
Why size stopped predicting usefulness
Stanford's 2025 AI Index found that the score gap between the top model and the tenth-best fell from 11.9% to 5.4% in a year, and that the top two were separated by just 0.7%. The same report found that the cost of running a system at GPT-3.5's level fell more than 280-fold between November 2022 and October 2024, driven by increasingly capable small models, while hardware costs fell about 30% a year.
Epoch AI, a research group that tracks these trends, found that the price of reaching a given level of performance fell by between 9 and 900 times a year, depending on the benchmark and the level chosen, though it cautions that the fastest drops may not last. In practice, last year's frontier becomes this year's budget tier, and a smaller model suited to your job often wins on cost. If data control is your worry, our guide to open-weight versus API models covers that trade-off.
04
What to measure instead
Cost per finished task. Independent testers at Artificial Analysis found that on coding-agent work, with each model running in its maker's own coding tool, Grok 4.5 cost $2.49 per task against $5.07 for GPT-5.5 and $11.80 for Claude Fable 5. Grok 4.5 also used far fewer tokens per task on average: 1.9 million, against 6.2 million for GPT-5.5 and 7.2 million for Fable 5. A model with a higher price per token can still be cheaper per job, and the reverse.
Reliability. On Artificial Analysis's knowledge test, Grok 4.5's accuracy rose from 35% to 52%, but its hallucination rate rose from 25% to 54%. A model can get more answers right and still make more things up, which is exactly what you need to catch before it writes to customers. Lian Jye Su of Omdia told InfoWorld that cost per job done remains the best way for most firms to judge an agent.
Fit. Size says nothing about whether a model handles your languages, file formats and tools. Meta's official list of 12 supported languages for Llama 4, for example, includes Hindi but not Malayalam. A smaller model that reads your invoices, answers customers in their language and connects to your CRM is worth more to you than a larger one that does not.
enterprise buyers should focus on cost per successful outcome rather than cost per token
05
What to do instead of reading spec sheets
You do not need a lab to compare models. You need a morning, a spreadsheet and real examples from your own business.
- Collect 20 real tasks: actual emails, documents and quote requests, each with the answer a good employee would give.
- Run every candidate on the same tasks with the same instructions, and record the cost and time per task.
- Score blind. Ask the person who owns the work to rate outputs without knowing which model wrote them.
- Count serious failures, not averages. One confident wrong price or deadline matters more than ten polished answers.
- Re-run the test quarterly. Releases now land roughly monthly, and last quarter's loser is often this quarter's best value.
Sources
Facts current as of 5 August 2026. Parameter figures for Grok models are Elon Musk's statements, not published specifications.