Reference
Models inside Kotia services
9 models from 4 companies. You pick a model for every message and can switch mid-conversation. If a model doesn't suit you, you can always choose another and find the one that performs best for your tasks.
Model comparison table
| Model | Context window | Cost | Files | Web search | Access |
|---|---|---|---|---|---|
Claude Haiku 5.5 default Anthropic · Fast, agent tasks | ≈ 1111 pages 1000000 tokens | low the cheapest one | Images, PDF, text | Built-in | Everyone |
GPT-6 Luna OpenAI · Fast, cheapest | ≈ 2042 pages 1050000 tokens | low ×1 vs the cheapest | Images, PDF, text | Built-in | Everyone |
Gemini 3.5 Flash-Lite Google · Editing, fast | ≈ 1944 pages 1000000 tokens | low ×3 vs the cheapest | Images, PDF, text | Built-in | Everyone |
DeepSeek V4 Flash DeepSeek · Cheap, text only | ≈ 1611 pages 1000000 tokens | medium ×4.4 vs the cheapest | Text only | Via Kotia | Everyone |
Gemini 3.6 Flash Google · Best for writing | ≈ 1944 pages 1000000 tokens | medium ×7.5 vs the cheapest | Images, PDF, text | Built-in | Subscription |
DeepSeek V4 Pro DeepSeek · Strong reasoning, text only | ≈ 1611 pages 1000000 tokens | high ×13.2 vs the cheapest | Text only | Via Kotia | Subscription |
Claude Sonnet 5.5 Anthropic · Balanced | ≈ 1111 pages 1000000 tokens | high ×20 vs the cheapest | Images, PDF, text | Built-in | Subscription |
GPT-6.1 Sol OpenAI · Strongest for documents | ≈ 2042 pages 1050000 tokens | high ×20 vs the cheapest | Images, PDF, text | Built-in | Subscription |
Claude Opus 5.5 Anthropic · Most capable | ≈ 1111 pages 1000000 tokens | high ×40 vs the cheapest | Images, PDF, text | Built-in | Subscription |
Sorted from the cheapest to the most expensive. A page is 1,800 characters.
What the terms above mean
- Context window
- How much text the model keeps in mind while replying: your question, the whole conversation, attached files and its own answer. Once the window fills up, older messages stop being taken into account. We show it in pages of plain text so it's easier to tell whether a document will fit.
- Page
- A rough unit of volume: 1,800 characters including spaces, about one A4 page in a regular font. The page count in the table is an estimate — the exact figure depends on the language of the text.
- Token
- A chunk of text that models count with: a word or part of a word. Providers measure the window and charge in tokens. As a rule of thumb, one word is one to three tokens.
- Cost
- How much is deducted from your balance for one reply. It depends on the model and on length: the longer the conversation and the files, the more each following reply costs. The table shows the tier and how many times more expensive the model is than the cheapest one.
- Files
- What you can attach to a message. «Images, PDF, text» — the model sees the image and the PDF pages. «Text only» — the model accepts TXT, MD, CSV and JSON and rejects an image or PDF right when you send it.
- Web search
- «Built-in» — the model searches the web itself, inside its reply. «Via Kotia» — the platform runs the search and opens the pages, then hands the findings to the model. The results are comparable; what differs is who searched and how it shows in the feed.
- Access
- «Everyone» — the model is available on any account. «Subscription» — only with an active subscription. There are no free models: tokens are deducted for every reply from any model; the subscription only decides which models you can choose.
- Default
- The model the chat runs on until you pick another one. You can switch at any point in the conversation — the next reply comes from the new model, previous ones stay as they are.
What makes each model different and which one to choose
Everything below comes from working with the models in Kotia, not from provider marketing.
| Model | What makes it different | When to choose it | When to avoid it |
|---|---|---|---|
Claude Haiku 5.5 default Anthropic · Fast, agent tasks |
| Agent-style work: canvas, tool chains, pipelines, digging through materials with search and source reading. Holds a long sequence of steps without losing the task. | Pure text work — style and editing are drier than Google's. Haiku 5.5 is cheap on short prompts; a turn with history runs on its pricier long-prompt tier. |
GPT-6 Luna OpenAI · Fast, cheapest |
| Long professional documents, analysis, structured texts. The junior model is the cheapest in the chat. | Precise tracking of conversation length — the provider offers no counter, so the remaining window is shown with a safety margin. |
Gemini 3.5 Flash-Lite Google · Editing, fast |
| Text: writing, editing, conversation, working with large volumes of material. The fastest replies in the chat. | Long agent chains with tools — on multi-step tasks it falls behind Anthropic. |
DeepSeek V4 Flash DeepSeek · Cheap, text only |
| Reasoning and problem analysis on a budget, when the work is text only. | Images and PDF (not accepted), confidential materials, urgent short questions. |
Gemini 3.6 Flash Google · Best for writing |
| Text: writing, editing, conversation, working with large volumes of material. The fastest replies in the chat. | Long agent chains with tools — on multi-step tasks it falls behind Anthropic. |
DeepSeek V4 Pro DeepSeek · Strong reasoning, text only |
| Reasoning and problem analysis on a budget, when the work is text only. | Images and PDF (not accepted), confidential materials, urgent short questions. |
Claude Sonnet 5.5 Anthropic · Balanced |
| Agent-style work: canvas, tool chains, pipelines, digging through materials with search and source reading. Holds a long sequence of steps without losing the task. | Pure text work — style and editing are drier than Google's. Haiku 5.5 is cheap on short prompts; a turn with history runs on its pricier long-prompt tier. |
GPT-6.1 Sol OpenAI · Strongest for documents |
| Long professional documents, analysis, structured texts. The junior model is the cheapest in the chat. | Precise tracking of conversation length — the provider offers no counter, so the remaining window is shown with a safety margin. |
Claude Opus 5.5 Anthropic · Most capable |
| Agent-style work: canvas, tool chains, pipelines, digging through materials with search and source reading. Holds a long sequence of steps without losing the task. | Pure text work — style and editing are drier than Google's. Haiku 5.5 is cheap on short prompts; a turn with history runs on its pricier long-prompt tier. |
Questions and answers
Where exactly do these models work?
In the chat and on the canvas — where you talk to the agent and build chains of steps. Article generation, research, news and AI detection run on their own models, and the choice in the chat does not affect them.
Are there free models?
No. Tokens are deducted for every reply from any model — free or purchased. A subscription doesn't make models free; it unlocks the strong models that are not on the list without it.
How do I switch models, and what happens to the conversation?
The model selector sits next to the input field and can be changed at any time. The next reply comes from the new model, and the whole conversation stays in place — the new model sees all of it.
The model didn't reply or returned an error. What happens?
Kotia makes up to three attempts on the same model. If that fails, you'll see an error and decide what to do: retry or switch to another model. Kotia never swaps the model silently — you always know who replied.
Why does the same question cost different amounts?
The price of a reply depends on length: the model reads not only the question but the whole conversation, attached files and search results. The longer the conversation, the more each following reply costs. A new chat for a new topic is the simplest way to pay less.
What if I run out of tokens in the middle of a reply?
The reply won't be cut off or lost: you'll receive it, and the balance may go slightly negative. The minus is cleared by your next top-up. To start a conversation you need a minimum balance.
Can I attach a PDF, an image or a spreadsheet?
Yes, with models marked «Images, PDF, text». A model marked «Text only» accepts TXT, MD, CSV and JSON and rejects an image or PDF right when you send it, rather than replying about an empty file.
What does «context window is full» mean?
The model can no longer fit the whole conversation, and older messages stop being taken into account. The usage meter next to the input field shows the remaining space with a margin. For a large document, pick a model with a wider window or start a new chat.
Where is the conversation stored, and does the model provider see it?
Conversations are stored in Kotia. With most models the provider doesn't retain them. The exception is DeepSeek: there the conversation stays with the provider and this cannot be disabled, so it's better not to send confidential materials there.
A stray monologue like «the user is asking…» appeared in the reply. Is that a bug?
That's the model's internal reasoning, which occasionally leaks into the visible reply — it happens with the junior DeepSeek model. The answer itself is correct. If it bothers you, send the message again or choose another model.