
Monthly Krater usage data: which AI models people choose, LLM market share by company and how both changed since our first report.
Among models people chose themselves on Krater in October 2026, Claude Sonnet 4.6 and Claude Sonnet 5.5 were the most used, each picked by about 13% of active users. By company, Anthropic led with 32.7% of chosen requests, ahead of OpenAI at 25.9% and Google at 19.7%. Compared with our September report, the biggest shifts are agent mode passing plain chat and a default model change that moved OpenAI to the top of total requests.

This is the second edition of our model usage report. The first, Most Used AI Models in 2026, covered the 90 days to September 6, 2026 and stays online unchanged, so every number here can be checked against it. From now on we publish one report per month, with shares only.
Our first report counted every chat and agent request, including the default model. To compare fairly, we ran that exact method again on October. Rerunning it on the original 90-day window reproduces the published company shares to within 0.1 points, so the two columns below are measured the same way.
| Company | June 8 to Sept 6, 2026 (first report) | October 2026 (same method) |
|---|---|---|
| OpenAI | 24.2% | 48.0% |
| Anthropic | 26.6% | 17.9% |
| 31.1% | 13.5% |
Read quickly, that looks like OpenAI winning the market in two months. It is not. In August we changed the default model in new chats from Gemini 2.5 Flash to GPT-5.6 Luna. Gemini 2.5 Flash was 19.8% of all requests in the first report and 2.6% in October. GPT-5.6 Luna alone made up 33.9% of all requests in October. Most of OpenAI's jump and Google's drop is the default changing hands, not people changing their minds.
That is the most useful lesson in the data so far: the default decides a large part of market share. It also explains why public "most used AI" rankings based on app traffic mostly measure which app people opened, not which model they would pick. For the rest of this report we leave both defaults out, and look only at what people chose.
Share of chat and agent requests where the user picked the model, by the company that built it. Both defaults (Gemini 2.5 Flash and GPT-5.6 Luna) and every Auto Router mode are excluded in every month, so the trend is like for like.
| Company | June | July | August | September | October |
|---|---|---|---|---|---|
| Anthropic | 38.9% | 41.7% | 35.5% | 34.2% | 32.7% |
| OpenAI | 19.3% | 22.9% | 21.0% | 23.8% | 25.9% |
| 21.3% | 16.9% | 14.1% | 13.8% | 19.7% | |
| DeepSeek | 1.8% | 1.6% | 9.2% | 8.8% | 8.4% |
| Z.ai | 0.4% | 0.2% | 0.4% | 2.4% | 3.6% |
| xAI | 8.7% | 4.6% | 4.7% | 4.3% | 2.3% |
| Perplexity | 2.8% | 2.5% | 2.6% | 2.6% | 2.3% |
Four trends hold up across the five months:
The first report could not show any of this, because it was a single 90-day window. Monthly editions make the direction visible.
Share of all chat and agent requests that ran in agent mode, where the AI plans steps, uses tools and works through a task.
| Month | Agent mode share |
|---|---|
| June 2026 | 23.3% |
| July 2026 | 45.7% |
| August 2026 | 49.5% |
| September 2026 | 50.5% |
| October 2026 | 54.0% |
The September report found agent requests at 76% of the volume of chat, about 43% of the total. Since September, agent mode is the majority. People increasingly ask the AI to finish a task, not just answer a question, and that changes which model matters: a model that handles tools and long tasks well now carries more of the work.
Ranked by the share of active users who chose each model at least once in October. Both defaults and the Auto Router are left out, and dated snapshots of the same model are counted together.
| Rank | Model | Company | Share of active users |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | Anthropic | 13.6% |
| 2 | Claude Sonnet 5.5 | Anthropic | 13.4% |
| 3 | DeepSeek V4.1 Flash | DeepSeek | 12.5% |
| 4 | Gemini 3.8 Flash | 12.2% | |
| 5 | GLM 5.3 Flash | Z.ai | 11.8% |
| 6 | Claude Sonnet 5 | Anthropic | 10.9% |
| 7 | Claude Opus 5.5 | Anthropic | 9.1% |
| 8 | GPT-6 Astra | OpenAI | 8.8% |
| 9 | GPT-5.6 Luna Pro | OpenAI | 8.6% |
| 10 | GPT-6.1 Sol | OpenAI | 7.9% |
Two patterns stand out. Fast, low-cost models take three of the top five spots: DeepSeek V4.1 Flash, Gemini 3.8 Flash and GLM 5.3 Flash. And people do not all move to the newest version: Claude Sonnet 4.6, 5 and 5.5 all sit in the top six at the same time.
Compared with the first report, the names at the top are almost entirely new. Its top ten included Claude Sonnet 4.5, Gemini 2.5 Pro and Grok 4.3, none of which is in October's top ten. Over a few months, model habits turn over quickly.
Share of active users who chose at least one model from each company in October. One person can count for several companies, so the numbers add up to more than 100%.
| Company | Share of active users |
|---|---|
| Anthropic | 45.1% |
| OpenAI | 35.6% |
| 22.7% | |
| DeepSeek | 18.8% |
| Z.ai | 13.6% |
| xAI | 8.8% |
| Perplexity | 7.0% |
| Mistral | 5.0% |
| Moonshot AI | 4.8% |
OpenAI's real reach is higher, because the default model is an OpenAI model and it is excluded here. 27% of active users chose at least one model from DeepSeek, Z.ai, Moonshot AI, Qwen, Meta or Mistral, so these companies are no longer a niche choice.
| Models used in October | Share of active users |
|---|---|
| 1 model | 36.1% |
| 2 to 3 models | 30.8% |
| 4 to 9 models | 24.5% |
| 10 or more models | 8.6% |
Almost two thirds of active users used two or more models, 57% used models from at least two companies and 33% from three or more. One more trend is worth watching: the share of requests on a model people picked themselves fell from 71% in July to 55% in October. Our data cannot say why. One likely reason is that newer users keep the default more often than people who have used Krater for months, which we will check in the next edition.
For quality scores rather than popularity, read our business benchmark hub. Popularity tells you what people use; the benchmark tells you what performs.
For a free AI consultation, Adam maps what is best for your team size and usage. Free, no strings attached.
Thirty minutes with Adam Johansson, an AI expert who maps out what actually fits your business: the right tools, plan and setup for your team size and usage. Free, no strings attached.
Free for any business, any size. We reply within one business day.

Adam Johansson
AI Advisor, Krater
Among models people chose themselves on Krater in October 2026, Claude Sonnet 4.6 (13.6% of active users) and Claude Sonnet 5.5 (13.4%) were the most used, followed by DeepSeek V4.1 Flash (12.5%). Counting every request including the default, GPT-5.6 Luna was used most, because it is the default in new chats.
Among requests where people chose the model, Anthropic led in October 2026 with 32.7%, followed by OpenAI at 25.9% and Google at 19.7%. Counting all requests including the default, OpenAI led with 48.0%.
Agent mode grew from about 43% of requests in our first report to 54% in October. OpenAI's total share doubled after a default model change, while among chosen models Anthropic's lead narrowed and OpenAI and DeepSeek gained.
Yes. Almost two thirds of active users used two or more models in October, and 57% used models from at least two AI companies.
Monthly. Each edition gets its own page, and earlier editions stay online unchanged so the data can be compared over time.