# AI models

*The article discusses artificial intelligence models on which you can build your own AI agents.*

The Selarti service provides 25 AI models: gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-5.4-pro, gpt-5.3-chat-latest, gpt-5.2, gpt-5.2-pro, gpt-5.1, gpt-5, gpt-5-mini, gpt-5-nano, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-o3, gpt-o3-deep-research, gpt-o4-mini, gpt-o4-mini-deep-research, gpt-o1, gpt-o3-mini, gpt-4o, gpt-4o-mini, gpt-4-turbo and gpt-4-turbo-preview. These models differ in functionality, performance and price, which allows you to choose the most suitable one for specific tasks.

At the moment we work with the response API, which allowed us to add all relevant models to the AI settings.

{% hint style="warning" %}
This resource is an open source; data uploaded to it becomes available to OpenAI. Encryption is absent.
{% endhint %}

You can review [the differences](https://platform.openai.com/docs/models/gpt-4o) of the models on the official OpenAI website, or you can read below in the article. You can also review Selarti pricing [here](/manual-en/financial-matters/plans-and-limits.md).

### Comparative table of AI models

Context window — the amount of text the model can view and process to generate a response. The larger it is, the more context from the dialogue the model will remember during a long conversation.

<table><thead><tr><th width="111.33331298828125">Name</th><th width="84.3333740234375">Input price</th><th width="83">Cached input price</th><th width="94.6666259765625">Output price</th><th width="100">Context window</th><th width="90">Speed</th><th width="148.3333740234375">Purpose</th><th width="97">Ability to reason</th><th>Additional model settings</th></tr></thead><tbody><tr><td>gpt-5.5</td><td>$6</td><td>$0.6</td><td>$36</td><td>-</td><td>Medium</td><td>Programming, analytics, document management, business tasks</td><td>Highest</td><td>Reasoning effort and verbosity</td></tr><tr><td>gpt-5.4</td><td>$3</td><td>$0.30</td><td>$18</td><td>1050000</td><td>Medium</td><td>Analytics, process automation, document handling</td><td>Highest</td><td>Reasoning effort and verbosity</td></tr><tr><td>gpt-5.4-mini</td><td>$0.9</td><td>$0.09</td><td>$5.4</td><td>400000</td><td>High</td><td>Customer support, CRM, assistants</td><td>Above average</td><td>Reasoning effort and verbosity</td></tr><tr><td>gpt-5.4-nano</td><td>$0.24</td><td>$0.024</td><td>$1.5</td><td>400000</td><td>Very high</td><td>FAQ, auto-replies, bulk processes</td><td>Medium</td><td>Reasoning effort and verbosity</td></tr><tr><td>gpt-5.4-pro</td><td>$36</td><td>-</td><td>$216</td><td>1050000</td><td>Slowest</td><td>Strategy, consulting, sophisticated analytics, system design</td><td>Highest</td><td>Reasoning effort and verbosity</td></tr><tr><td>gpt-5.3-chat-latest</td><td>$2.10</td><td>$0.21</td><td>$16.80</td><td>128000</td><td>Medium</td><td>Customer Support, Chatbots, Assistants, Communications</td><td>High</td><td>Reasoning effort</td></tr><tr><td>gpt-5.2</td><td>$2.10</td><td>$0.21</td><td>$16.80</td><td>400000</td><td>Average</td><td>Development, testing, research</td><td>High</td><td>Effort of reasoning and verbosity</td></tr><tr><td>gpt-5.2-pro</td><td>$25.20</td><td>-</td><td>$201.60</td><td>400000</td><td>Slow</td><td>Content, analytics, texts</td><td>High</td><td>Effort of reasoning and verbosity</td></tr><tr><td>gpt-5.1</td><td>$1.50</td><td>$0.15</td><td>$12.00</td><td>400000</td><td>Average</td><td>Development, testing, research</td><td>Above average</td><td>Effort of reasoning and verbosity</td></tr><tr><td>gpt-5</td><td>$1.50</td><td>$0.15</td><td>$12.00</td><td>400000</td><td>Medium</td><td>Sales, training, consulting, strategy</td><td>High</td><td>Effort of reasoning and verbosity</td></tr><tr><td>gpt-5-mini</td><td>$0.30</td><td>$0.03</td><td>$2.40</td><td>400000</td><td>Above average</td><td>Customer support, CRM, assistants</td><td>Above average</td><td>Effort of reasoning and verbosity</td></tr><tr><td>gpt-5-nano</td><td>$0.06</td><td>$0.006</td><td>$0.48</td><td>400000</td><td>High</td><td>FAQ, autoresponders, mass processes</td><td>Medium</td><td>Effort of reasoning and verbosity</td></tr><tr><td>gpt-4.1</td><td>$2.40</td><td>$0.6</td><td>$9.60</td><td>1047576</td><td>Medium</td><td>Content, analytics, texts</td><td>Above average</td><td>Creativity and variety</td></tr><tr><td>gpt-4.1-mini</td><td>$0.48</td><td>$0.12</td><td>$1.92</td><td>1047576</td><td>Above average</td><td>General tasks, support</td><td>Medium</td><td>Creativity and variety</td></tr><tr><td>gpt-4.1-nano</td><td>$0.12</td><td>$0.03</td><td>$0.48</td><td>1047576</td><td>Very high</td><td>Simple answers, notifications</td><td>Below average</td><td>Creativity and variety</td></tr><tr><td>gpt-o3</td><td>$2.40</td><td>$0.6</td><td>$9.6</td><td>200000</td><td>Very high</td><td>Simple answers, support</td><td>Highest</td><td>Effort of reasoning</td></tr><tr><td>gpt-o3-deep-research</td><td>$12.00</td><td>$3</td><td>$48.00</td><td>200000</td><td>Slow</td><td>Analytics, data processing</td><td>Highest</td><td>Creativity and variety</td></tr><tr><td>gpt-o4-mini</td><td>$1.32</td><td>$0.33</td><td>$5.28</td><td>200000</td><td>Medium</td><td>Everyday tasks, texts</td><td>Above average</td><td>Effort of reasoning</td></tr><tr><td>gpt-o4-mini-deep-research</td><td>$2.40</td><td>$0.66</td><td>$9.60</td><td>200000</td><td>Medium</td><td>Analysis, summarization, reports</td><td>Above average</td><td>Creativity and variety</td></tr><tr><td>gpt-o1</td><td>$18.00</td><td>$9</td><td>$72.00</td><td>200000</td><td>Slow</td><td>Tests, demonstrations</td><td>Above average</td><td>Depth of thought</td></tr><tr><td>gpt-o3-mini</td><td>$1.32</td><td>$0.66</td><td>$5.28</td><td>200000</td><td>Medium</td><td>For logical tasks, programming, data work, complex assistants</td><td>Above average</td><td>Depth of thought</td></tr><tr><td>gpt-4o</td><td>$3.00</td><td>$1.5</td><td>$12.00</td><td>128 000</td><td>Medium</td><td>Working with text and images</td><td>Medium</td><td>Creativity and variety</td></tr><tr><td>gpt-4o-mini</td><td>$0.18</td><td>$0.09</td><td>$0.72</td><td>128000</td><td>Above average</td><td>Visual tasks, fast responses</td><td>Below average</td><td>Creativity and variety</td></tr><tr><td>gpt-4-turbo</td><td>$12.00</td><td>-</td><td>$36.00</td><td>128000</td><td>Medium</td><td>General tasks, content, support</td><td>Below average</td><td>Creativity and variety</td></tr><tr><td>gpt-4-turbo-preview</td><td>$12.00</td><td>-</td><td>$36.00</td><td>No information yet</td><td>No information yet</td><td>New features, tests</td><td>No information yet</td><td>Creativity and variety</td></tr></tbody></table>

Input and output prices are indicated per 1 million tokens.

The table does not contain all information about models available for early access — the capabilities information is still being gathered.

Below we describe the features of the models in a bit more detail, and at the end of the article we share our experience and observations about creating the Ilyusha bot on the o3 and GPT-4o models.

### gpt-5.5

OpenAI’s flagship reasoning model for complex intellectual tasks, programming, AI agents, and analytics. It is capable not only of answering questions, but also of independently managing complex multi-step workflows. The model is better at maintaining long reasoning chains, planning steps, checking for errors, handling uncertainty, and working with large context windows.

Recommendations:\
Choose this model if your AI agent needs to solve deep reasoning tasks, process and generate complex code, operate autonomously, or work with large contextual information.

Use cases:

* Complex programming.
* Advanced AI agents.
* In-depth research.
* Data analysis

Not recommended for:\
Simple high-volume tasks that require maximum speed and minimal cost (for example, basic chatbots).

### gpt-5.4

A model with enhanced logic, accuracy and understanding of complex tasks. It copes better with multi-step scenarios, structuring information and working with complex instructions.
\
Combines the strengths of reasoning and communication.

Recommendations:
\
Choose if you need a universal AI that is equally good at dialogue and solving complex problems.

Examples of use:

* Advanced chatbots with scripting logic and branching.
* Analytical assistants for business and reporting.
* AI assistants for working with documents, tables, regulations, knowledge bases.
* Automation of internal processes and communications.

Not recommended:
\
For the deepest expert analysis — it is better to use the pro version in such tasks.

### gpt-5.4-mini

A compact and fast version of GPT-5.4. It operates more than twice as fast as GPT-5 mini and delivers performance close to the full GPT-5.4 model on a range of benchmarks. It is also more affordable and cost-efficient than the flagship version.

**Recommendations:**\
Well suited for tasks where speed, cost efficiency, and stable quality are important.

**Use cases:**

* Programming and code autocompletion
* Workflow automation
* Document analysis and long-context processing
* Large-scale content generation

**Not recommended for:**\
Tasks where maximum reasoning depth, precision, or complex planning is critical.

### gpt-5.4-nano

The most compact and affordable version of GPT-5.4.

**Recommendations:**\
Best suited for ultra-fast, high-volume tasks with minimal latency.

**Use cases:**

* Message classification (spam/not spam, complaint/question/order)
* Processing customer support requests
* Data extraction (emails, phone numbers, dates, amounts from text)
* Content moderation (fast screening and filtering)

**Not recommended for:**\
Tasks that require deep reasoning, high precision, or handling complex context.

### gpt-5.4-pro

Professional version of gpt-5.4 with maximum depth of thinking, accuracy and the ability to work with complex systems. Best understands hidden dependencies, complex contexts and non-standard tasks.
\
Gives the most reasonable, structured and thoughtful answers.

Recommendations:
\
Choose if you need an AI expert for complex, critical or strategic tasks.

Examples of use:

* Strategic analysis, consulting, decision-making.
* Design architectures, systems, and processes.
* Deep data analytics and business planning.
* Complex automation and integration scenarios.

Not recommended:
\
For simple tasks, quick answers and mass scenarios, the model will be redundant in capabilities and expensive.

### gpt-5.3-chat-latest

Chat version of the gpt-5.3 line, which focuses on live dialogue, naturalness and user adaptation. He better understands the context of communication, tone, intentions and style, reacts faster and conducts a conversation more «humanly».
\
Optimized for real communication, support and everyday tasks.

Recommendations:
\
Choose if you need the most pleasant, fast and natural dialogue with the user.

Examples of use:

* AI assistant in support chat and client service.
* Chatbots for sites, social networks, instant messengers.
* Personal assistants (planning, answers, tips).
* Content assistants for texts, letters, messages.

Not recommended:
\
For complex analytics, strategies and tasks with long chains of logic — in these cases, models with an emphasis on logic and accuracy are better suited.

### gpt-5.2

An updated version of gpt-5 with improved reasoning and resilience in long dialogues. Holds context better, less often "loses the thread" of the conversation, works more carefully with logic, instructions and complex scenarios.\
Fewer errors in multi-step tasks, more stable in business processes.

Recommendations:\
Choose it if you need the most reliable AI for complex dialogues and scenarios where consistency and accuracy are important.

Examples of use:

* AI assistant for customer support with a long history of interactions.
* Business bot for managing deals and projects.
* Analytical assistant for reports, planning and strategies.
* AI operator working with regulations and instructions.

Not recommended:\
For simple and short tasks — excessive in capability and cost.

### gpt-5.2-pro

Professional version gpt-5.2 with maximum depth of analysis, logic and understanding of user intent. Works excellently with abstract tasks, complex systems, creativity and strategy.\
Best among the models in the lineup at explaining decisions and justifying conclusions.

Recommendations:\
Choose it if the AI must think "like an expert" and be a full intellectual partner.

Examples of use:

* AI strategist, business consultant, analyst.
* Complex AI coaches and mentors.
* Generation of scenarios, concepts, product architectures.
* Premium AI characters with high awareness and speech style.

Not recommended:\
For mass, fast or template responses — too powerful and expensive.

### gpt-5.1

An intermediate version between gpt-5 and gpt-5.2. Slightly weaker in deep analysis and long chains of reasoning, but still a very "smart" and flexible model.\
Well suited for dialogues where a balance between quality and speed is important.

Recommendations:\
An excellent option if you need a strong AI without overpaying for maximum power.

Examples of use:

* Online consultants and support assistants.
* AI managers for sales and customer support.
* Content assistants, editors, marketer helpers.
* Universal chatbots for websites and apps.

Not recommended:\
For ultra-critical analytical tasks or expert consultations — better to take gpt-5.2 or gpt-5.2-pro.

### gpt-5

OpenAI's flagship model of 2025. Understands complex contexts, emotions, intentions, and subtexts. Can reason, joke, remember communication style, and adapt to a brand. Excellent for AI characters that should behave like real employees.

Recommendations:\
Choose this if you need an intelligent, «human» AI that not only answers but thinks, analyzes and speaks deliberately.

Examples of use:

* AI sales manager capable of negotiating and handling objections.
* HR bot conducting interviews and evaluating candidates' answers.
* Premium assistant supporting personalized dialogues with customers.
* AI coach, consultant, scriptwriter, analyst.

Not recommended:\
If speed and low cost are required for simple tasks (for example, an FAQ bot or autoresponder).

### gpt-5-mini

A lightweight version of gpt-5. Retains almost all the intelligence of the senior model, but pays slightly less attention to nuances and subtle emotional moments. Results are still high-level, but responses are shorter and faster.

Recommendations:\
An optimal option for most companies — a balance of realism, speed and price.

Examples of use:

* Online store online consultant.
* Assistant for selecting tariffs and services.
* Customer success manager.
* AI assistant for CRM or website.

Not recommended:\
For ultra-creative or analytical scenarios — better to choose gpt-5 or gpt-4.1.

### gpt-5-nano

Economy version of gpt-5, designed for short, quick and simple responses. Does not reason deeply, but handles template tasks well.

Recommendations:\
For mass bots, autoresponders and 24/7 support.

Examples:

* Autoresponder: We have received your request», «An operator will reply shortly»
* FAQ bot: «How do I place an order?», «How do I return a product?»
* AI agent for notifications and short prompts.

Not recommended:\
If the bot must converse like a human or provide detailed answers.

### gpt-4.1

A reliable model capable of generating texts, scripts, ideas, and analyzing data. Best suited for creative and content tasks.

Recommendations:\
For projects where the AI must write beautifully, logically and interestingly.

Examples:

* AI copywriter for social networks, blogs, newsletters.
* Scriptwriter or author of educational materials.
* Analyst creating reports, summaries, conclusions.
* Assistant for marketing or content department.

Not recommended:\
If a conversational style and emotional adaptation are required (use gpt-5 for that).

### gpt-4.1-mini

A fast version of gpt-4.1 — a universal «workhorse». Can write texts, answer questions, give advice, but without deep analytics.

Recommendations:\
A great choice for small companies that need a reliable AI without frills.

Examples:

* Customer support.
* Product and service descriptions.
* Maintaining a knowledge base.
* Processing requests and messages.

### gpt-4.1-nano

Responds briefly and directly. Does not perform complex reasoning, but is inexpensive and works quickly.

Examples:

* Chatbot for quick answers.
* Auto-notifications and short instructions.
* Technical answers without creativity.

Recommendations:\
For internal automation, help systems, FAQ sections.

### gpt-03

A simplified model without emotions but very stable. Ideal for simple texts and standard responses.

Examples:

* Simple tech support.
* Creating short instructions.
* Answers to template questions.

Recommendations:\
For minimal AI tasks where stability matters more than intelligence.

### gpt-03-deep-research

Specialized for analyzing large texts and finding patterns. Performs comparisons, draws conclusions, summaries and analyses.

Examples:

* Analysis of customer reviews.
* Competitor comparisons.
* Searching for semantic patterns in data.
* Creating reports.

Recommendations:\
For analytical AI employees, marketers and researchers.

### gpt-04-mini

A reliable, fast and economical model. Writes texts well, but does not reason deeply.

Examples:

* Generation of short descriptions.
* Assistant for routine tasks (emails, notifications).
* Simple chatbots.

Recommendations:\
Suitable for starting a project or tests.

### gpt-04-mini-deep-research

Able to analyze texts, make generalizations and draw conclusions.

Examples:

* Preparing reports from reviews.
* Analysis of correspondence.
* Summarizing data from CRM.

### gpt-01

Used only for demonstrations or compatibility tests. Not suitable for production AI employees.

### gpt-03-mini

A fast and budget version for simple tasks.

Examples:

* Feedback forms.
* Chatbot with canned responses.
* Short template responses to users.

### gpt-4o

Understands not only text but also images. Can describe, analyze and compare visual information.

Examples:

* Analysis of product photos or documents (image processing is not supported in Selarti).
* Assistant checking design or mockups.
* AI assistant for visual content.

Recommendations:\
Ideal for companies working with graphics, products, documents.

### gpt-4o-mini

A fast and lightweight multimodal model.

Examples:

* Assistant for visual content for marketing (image processing is not supported in Selarti).
* Automatic image checking (image processing is not supported in Selarti).
* Recognition of screenshots, receipts, forms (image processing is not supported in Selarti).

### gpt-4-turbo

An optimized version of gpt-4 — combines intelligence, speed and affordability. Suitable for almost any type of AI employee.

Examples:

* Request manager.
* Assistant in CRM.
* Content bot that trains and informs customers.
* Virtual consultant in a messenger.

Recommendations:\
If you are unsure which model to choose — start with gpt-4-turbo.

### gpt-4-turbo-preview

The same turbo architecture but with new capabilities (improved context, speech stylization, memory). Sometimes it can be unstable but it delivers the newest features earlier than others.

Examples:

* Testing new AI employees.
* Prototypes where flexible style and natural speech are important.

### Our experience creating a bot for technical support on the o3-mini and gpt-4o models

For several months now we have been creating a bot for our onlinePBX project. This bot will answer customers' simple questions and escalate to staff if the question is complex. By the way, we named it Ilya; since it is a technical support employee, it should have a name.

In the process of creating the bot we used two models — o3-mini and gpt-4o. Below we share our observations.

#### Working with documentation:

* o3-mini refers to documentation even if instructed not to. It usually does this in a citation format that displays incorrectly. Sometimes it replaces quotation marks with emojis.
* o3-mini searches for mentions of the topic across all files and does this more thoroughly than gpt-4o.
* gpt-4o also refers to documentation; the prohibition does not always work.
* gpt-4o mentions uploaded files if they contain no information, for example it may say: «There is no such information in the uploaded files».
* Formatting the knowledge base in both cases requires a good structure — it is important to highlight items and subitems with headings so that models can save time and resources by «reading» only what relates to the client's specific question.

#### Model thinking:

* If there is no clear answer to a customer's question in the documentation, gpt-4o may give a very vague answer or say that such information does not exist. In the same situation o3 will analyze the information and invent an answer. Sometimes it will be correct, and other times completely made up.
* o3-mini has a larger context window, so it will process more information than gpt-4o.
* gpt-4o provides an answer quickly, typically within 10 seconds.
* o3-mini responds longer, depending on the thinking parameter set — about 15 seconds at minimum, 20–30 seconds at medium, and about a minute at high level.

#### Transferring the question to an operator:

* Both o3-mini and gpt-4o require clear instructions for handing the dialogue over to an operator, but for o3-mini this must be described in much more detail because the model will assume it can solve the issue itself and will invent answers that do not match reality.
* Unlike o3-mini, gpt-4o can hand the dialogue to an operator if it lacks information to answer. o3-mini does not understand the concept of lack of information and continues to fabricate answers.

#### Text formatting:

* gpt-4o uses Markdown formatting, which includes bold, italics, and link formatting as hyperlinks rather than raw URLs for better readability. If the bot will be used through a helpdesk, all markup should be disabled, otherwise text emphasis and links will display incorrectly.
* o3-mini produces text in a simple structure, can split it into items, but will send links as raw URLs.

#### Other features:

* When composing a prompt (instruction) gpt-4o easily works by roles and plays a character, can use more «friendly» formulations and convincingly pretend to be a support staff member.
* o3-mini does poorly with roleplay; if asked to be more emotional and friendly it is more likely to use expressive punctuation, for example adding more exclamation marks. o3-mini is more like a dry reference manual than an empathetic employee.


---

# Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://manual.selarti.com/manual-en/additional/ai-models.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.