LLM Statistics 2026: Key Numbers, Data & Facts

Large language models are the heart of the AI revolution. But how many are there, really? Who builds them? What do they cost? And which model is actually the best?

The short answer:

It has gotten messy. In 2026, a new top-tier model shows up roughly every month, prices swing by a factor of 600, and the single most important metric of the past few years, the parameter count, is something the big labs no longer disclose at all.

In this article, I sort through the numbers. API models come from our centrally maintained LLM database, the same one that powers tools like the API cost calculator. Open weights live in the separately generated open source directory. Both sources reflect the state of September 2026.

1. How Many Large Language Models Are There in 2026?

Our database currently tracks 288 large language models from 35 different providers, from GPT-2 in 2019 to the latest flagships in September 2026, including GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1, Mythos 5.1, and Qwen 3.8 Flash. Qwen 3.8 Flash Next is listed separately in the open source directory so the open weights are not counted as a second API model. This is deliberately a curated selection of the most important models, not a claim to completeness.

For context:

According to the Stanford AI Index 2026, US labs alone shipped around 60 notable models in 2025, Chinese providers about 35. More than 90% of all significant frontier models now come from industry rather than academic research. The market has professionalized and concentrated.

2. The Biggest LLM Providers by Model Count

A simple indicator of how active a lab is: the number of models it maintains. The chart below shows how many of the models we track belong to each provider:

Source: gradually.ai LLM database

PNG|SVG

CC BY 4.0

gradually.ai

Alibaba leads with 43 models, followed by OpenAI with 41 and Google with 29. That number only measures how deeply a lab maintains its lineup, though, not actual usage. Real market share looks different: in AI chatbot web traffic, ChatGPT dominates, while Gemini and Claude follow behind.

3. Parameters and Architecture: The End of Size Disclosures

For years, the parameter count was the most important metric for a model. GPT-3 had 175 billion, GPT-4 an estimated 1.76 trillion. Then the labs stopped reporting the number.

Today the rule is:

For every current frontier model from OpenAI, Anthropic, Google, and xAI, the parameter count is officially unknown. Model size has become a trade secret. Concrete, confirmed numbers almost only exist for open-weights models, and those are huge. The new leader is Moonshot AI's Kimi K3 at an official 2.8 trillion parameters (unveiled July 16, 2026, with weights released July 27 under Moonshot’s own “Kimi K3 License”):

Kimi K3 MoE, 896 experts (16 active)

2.8T

Qwen 3.8 Max 0902 MoE, 95B active

2.4T

DeepSeek-V4-Pro MoE, 49B active

1.6T

Kimi K2.6 MoE, 32B active

1T

GLM-5.2 MoE, 40B active

744B

DeepSeek V3.2 MoE, 37B active

685B

Mistral Large 3 MoE, 41B active

675B

DeepSeek-V4.1-Flash MoE, 8B active (input)

552B

Llama 4 Maverick MoE, 17B active

400B

Grok-1 MoE (2024)

314B

Source: gradually.ai LLM database

PNG|SVG

CC BY 4.0

gradually.ai

The architecture is the striking part. Almost all large models today use a Mixture-of-Experts (MoE) design, where only a fraction of the parameters is active per request. DeepSeek-V4-Pro has 1.6 trillion parameters but activates only 49 billion per token, around 3%. Qwen 3.8 Flash Next combines 125 billion main-model parameters with 51 billion N-gram embedding parameters and activates 6 billion parameters per token. That makes giant models affordable to run. In total, 109 of the tracked models are built as MoE.

You can filter and search the full parameter database by provider, size, and type below. For most current frontier models, the parameter column deliberately reads "unknown":

4. Context Windows: From 200,000 to 10 Million Tokens

The context window determines how much text a model can process at once. Here the orders of magnitude have multiplied over the past two years. The overview below covers more than 320 current models, sortable and filterable by provider: