> ## Documentation Index
> Fetch the complete documentation index at: https://tinytalk.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Models

> Choose from 50+ AI models across multiple providers.

Tiny Talk gives you access to models from multiple **model providers** (OpenAI, Anthropic, Google, Meta, xAI, DeepSeek, Mistral, Cohere, and Qwen) served through different **inference providers** (OpenAI, OpenRouter, Groq, and Azure). Each model has different strengths, speeds, and credit costs.

## How credits work

On plans that include AI credits (Basic AI, Standard AI, Pro AI), each agent response consumes credits based on the model used. Credits reset monthly.

| Plan        | Monthly Credits |
| ----------- | --------------- |
| Basic AI    | 2,000           |
| Standard AI | 10,000          |
| Pro AI      | 40,000          |

Most models cost **1 credit** per response. Mid-tier models (GPT-4o, Claude Sonnet, Gemini Pro) cost **2 credits**. Higher-tier models (GPT-4, o1) cost **5 credits**, and top-tier models (Claude Opus 4) cost **10 credits**.

## Bring Your Own Key (BYOK)

On legacy plans (Basic, Team, Enterprise) and some other plans, you can use your own API key instead of credits. Go to **Integrations → Hub** and enter your inference provider API key. The agent will use your key for all API calls, with no credit consumption on Tiny Talk's side.

After entering a key, click **Verify** to test it. Tiny Talk will check the key and display the models available for it. Make sure the model you want to use for your agent appears in this list.

<Warning>
  An **OpenAI key is always required** on BYOK plans — Tiny Talk uses OpenAI's embedding model for knowledge base indexing and search, regardless of which model you choose for agent responses.
</Warning>

<Info>
  Using your own OpenAI key? See [OpenAI's pricing page](https://developers.openai.com/api/docs/pricing) for per-token costs of each model.
</Info>

## Available models

### OpenAI

| Model         | Credit Cost | Reasoning         | Features                                          |
| ------------- | ----------- | ----------------- | ------------------------------------------------- |
| GPT-3.5 Turbo | 1           | —                 | Fast, affordable general-purpose model            |
| GPT-4         | 5           | —                 | High-intelligence model for complex tasks         |
| GPT-4 Turbo   | 5           | —                 | Faster, cheaper GPT-4 with vision support         |
| GPT-4o        | 2           | —                 | Versatile flagship model with vision              |
| GPT-4o Mini   | 1           | —                 | Fast, affordable with vision support              |
| GPT-4.1       | 2           | —                 | Flagship for complex tasks                        |
| GPT-4.1 Mini  | 1           | —                 | Balanced intelligence, speed, and cost            |
| GPT-4.1 Nano  | 1           | —                 | Fastest, most cost-effective GPT-4.1              |
| GPT-5         | 1           | Yes               | Flagship for coding, reasoning, and agentic tasks |
| GPT-5 Mini    | 1           | Yes               | Faster, cost-efficient version of GPT-5           |
| GPT-5 Nano    | 1           | Yes               | Fastest, cheapest GPT-5 variant                   |
| GPT-5.1       | 2           | Yes               | Advanced reasoning with vision                    |
| GPT-5.2       | 2           | Yes               | Advanced reasoning with vision                    |
| GPT-5.2 Chat  | 1           | Yes (medium only) | Cost-efficient reasoning with vision              |
| o1            | 5           | Yes               | Reasoning model with chain-of-thought             |
| o3-mini       | 1           | Yes               | Fast reasoning model                              |

### Anthropic (via OpenRouter)

| Model             | Credit Cost | Reasoning | Features                                         |
| ----------------- | ----------- | --------- | ------------------------------------------------ |
| Claude Opus 4     | 10          | —         | Top coding model with sustained performance      |
| Claude Sonnet 4.6 | 2           | —         | Latest Sonnet with enhanced coding and reasoning |
| Claude Sonnet 4   | 2           | —         | Enhanced coding and reasoning with precision     |
| Claude 3.7 Sonnet | 2           | —         | Improved reasoning and problem-solving           |
| Claude 3 Haiku    | 1           | —         | Fastest Anthropic model, near-instant responses  |

### Google (via OpenRouter)

| Model                  | Credit Cost | Reasoning | Features                                         |
| ---------------------- | ----------- | --------- | ------------------------------------------------ |
| Gemini 3.1 Pro Preview | 2           | Yes       | Latest flagship with reasoning, 1M token context |
| Gemini 3 Flash Preview | 1           | Yes       | Optimized for speed, 1M token context            |
| Gemini 2.5 Pro         | 2           | —         | Advanced reasoning, coding, and math             |
| Gemini 2.5 Flash       | 1           | —         | Fast workhorse for reasoning tasks               |
| Gemma 3 27B            | 1           | —         | Multilingual, 140+ languages                     |

### Meta (via OpenRouter)

| Model                  | Credit Cost | Reasoning | Features                           |
| ---------------------- | ----------- | --------- | ---------------------------------- |
| Llama 4 Maverick       | 1           | —         | Multimodal, 12 languages           |
| Llama 4 Scout          | 1           | —         | MoE model for multilingual tasks   |
| Llama 3.3 70B Instruct | 1           | —         | Multilingual dialogue, 8 languages |
| Llama 3.1 8B Instruct  | 1           | —         | Lightweight, low-latency           |

### xAI (via OpenRouter)

| Model            | Credit Cost | Reasoning | Features                                    |
| ---------------- | ----------- | --------- | ------------------------------------------- |
| Grok 4.1 Fast    | 1           | Yes       | Fast reasoning model, 256k context          |
| Grok 4           | 2           | Yes       | Latest reasoning model, 256k context        |
| Grok 3 Beta      | 2           | —         | Enterprise use cases, deep domain knowledge |
| Grok 3 Mini Beta | 1           | —         | Lightweight thinking model                  |

### Other model providers (via OpenRouter)

| Model             | Credit Cost | Reasoning | Model Provider |
| ----------------- | ----------- | --------- | -------------- |
| DeepSeek V4 Pro   | 1           | Yes       | DeepSeek       |
| DeepSeek V4 Flash | 1           | Yes       | DeepSeek       |
| DeepSeek-V3.2     | 1           | Yes       | DeepSeek       |
| DeepSeek-R1       | 1           | Yes       | DeepSeek       |
| DeepSeek V3 0324  | 1           | —         | DeepSeek       |
| Mistral Nemo      | 1           | —         | MistralAI      |
| Mistral Small 3   | 1           | —         | MistralAI      |
| Command R         | 1           | —         | Cohere         |
| Qwen3 235B A22B   | 1           | —         | Qwen           |
| Qwen 2.5          | 1           | —         | Qwen           |

### Groq (inference provider)

| Model                | Credit Cost | Reasoning             | Features                              |
| -------------------- | ----------- | --------------------- | ------------------------------------- |
| Llama 3.3 70B        | 1           | —                     | Ultra-fast inference on Groq hardware |
| Llama 3.1 8B Instant | 1           | —                     | Instant responses, low latency        |
| GPT-OSS 120B         | 1           | Yes (low/medium/high) | Large reasoning model                 |
| GPT-OSS 20B          | 1           | Yes (low/medium/high) | Smaller reasoning model               |

### Azure OpenAI (inference provider, EU data residency)

Available on **Pro** plan only. These models run on Azure's European (Germany) infrastructure for data residency compliance.

| Model                     | Credit Cost | Reasoning | Features                |
| ------------------------- | ----------- | --------- | ----------------------- |
| GPT-4o (Azure EU/DE)      | 2           | —         | European data residency |
| GPT-4o Mini (Azure EU/DE) | 1           | —         | European data residency |

<Tip>
  Need a different model on Azure EU? Pro plan users can contact support to request provisioning of additional models.
</Tip>

## Reasoning-capable models

Reasoning-capable models spend additional compute "thinking" before they answer, which improves quality on complex, multi-step questions in exchange for higher latency and credit cost. When you select one of these models, the **Reasoning Effort** and **Reasoning Summary** controls appear in your agent's model configuration (and the temperature slider is hidden). See [Reasoning configuration](/features/agent-settings#reasoning-configuration) for details.

Models marked **Yes** in the **Reasoning** column of the tables above support these controls. They span several providers:

* **OpenAI** — o1, o3-mini, GPT-5, GPT-5 Mini, GPT-5 Nano, GPT-5.1, GPT-5.2, GPT-5.2 Chat
* **Google** — Gemini 3.1 Pro Preview, Gemini 3 Flash Preview
* **xAI** — Grok 4.1 Fast, Grok 4
* **DeepSeek** — DeepSeek V4 Pro, DeepSeek V4 Flash, DeepSeek-V3.2, DeepSeek-R1
* **Groq** — GPT-OSS 120B, GPT-OSS 20B

<Info>
  Available effort levels vary by model — for example, **GPT-5.2 Chat** only supports **Medium**, and **GPT-OSS 120B/20B** only support **Low**, **Medium**, and **High**. The dashboard only shows values the selected model accepts.
</Info>

## Selecting a model

Go to your agent's **Settings** and choose a model from the dropdown. Consider:

* **Cost** — 1-credit models are most efficient for high-volume agents
* **Quality** — Premium models (GPT-5, Claude Sonnet/Opus) produce better responses for complex queries
* **Speed** — Groq models offer the fastest inference; Mini/Nano variants are faster than full models
* **Reasoning** — [Reasoning-capable models](#reasoning-capable-models) deliver better answers on complex questions at the cost of latency and extra credits
* **Data residency** — Use Azure EU models if you need European data processing (Pro plan)