> For the complete documentation index, see [llms.txt](https://docs.gallabox.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.gallabox.com/chat-ai-agent/choosing-an-ai-model-for-your-ai-agent.md). # Choosing an AI Model for Your AI Agent ## When you build an AI Agent, you can choose which AI model powers it. Different models balance **cost (credit usage)**, **accuracy**, and **speed** differently. This guide helps you pick the right one. **The short version:** Start with a **balanced** model. Move to a **lightweight** model if you want to save credits on simple tasks, or an **advanced** model if your agent handles complex reasoning. Then test with your own knowledge base and adjust.
### How model choice affects credits usage AI Credits are used based on the number of **tokens** processed — both the text sent to the model (**input**) and the text it generates (**output**). **What is a token?** A token is a small chunk of text a model reads or writes — roughly **4 characters, or about ¾ of a word**. Two things drive how many credits a model uses: 1. **The model's rate.** More advanced models charge more credits per 1,000 tokens. Output usually costs several times more than input. 2. **How much the model "thinks."** Advanced models can reason through a problem before answering. Better reasoning improves accuracy on hard questions but generates more output tokens — which uses more credits. **Rule of thumb:** the smarter the model and the harder you make it think, the more credits each reply uses. ### Model tiers at a glance Models fall into three broad tiers. Use this to narrow down quickly:
TierModelsCredit usageBest accuracy forGood forSpeed
Lightweight

  • DeepSeek-V4-Flash
  • gpt-4o-mini
  • gpt-5.4-nano
  • gpt-oss-120b
  • gpt-5-mini
LowSimple, well-defined tasksFAQs, routing, short answers, high message volumes⚡ Fastest
Balanced

  • gemini-2.5-flash
  • Qwen3.6 Plus
  • Kimi K2.5
  • gpt-5.4-mini
  • Kimi K2.6
  • Grok 4.3
MediumMost everyday support & salesGeneral chatbots, KB Q&A, lead qualificationFast
Advanced

  • gpt-5
  • gemini-2.5-pro
  • gpt-4o
  • gpt-5.4
 HighComplex reasoning & nuanceMulti-step reasoning, sensitive/complex queriesSlower
**About "thinking" options** Some models — including `gpt-5`, `gpt-5.4`, `gemini-2.5-pro`, and `Grok 4.3` — let you set a **thinking level** (`none`, `low`, `medium`, or `high`). A higher thinking level makes the model reason more before answering, which can improve accuracy on hard questions but generates more tokens, so it **uses more credits**. `Grok 4.3` appears in the dropdown both on its own and as ready-made `low` / `medium` / `high` thinking presets. #### Image recognition needs a vision-capable model If you want your agent to **understand images** customers send — for example a photo of a product, a screenshot, or a document — enable **Process Images** in the agent settings **and** choose a model that supports vision. Not every model can "see" images; sending an image to a text-only model won't work. `NOTE: processing an image consumes credits, and images generally use more input tokens than text` **Models that support image recognition** * `gpt-4o`, `gpt-4o-mini` * `gpt-5`, `gpt-5-mini` * `gpt-5.4`, `gpt-5.4-mini` * `gemini-2.5-flash`, `gemini-2.5-pro` * `Grok 4.3` If your agent uses Tools or Actions, for example checking an order status, booking a slot, calling an external API, or extracting structured data, the model has to reliably decide when to call a tool, which one, and how to fill in the details. This is harder than plain chat, and lightweight models are noticeably less reliable at it. **We recommend a Balanced or Advanced model for any agent that uses Tools or Actions.** ### How much do models actually use? (example) Credit use depends on your prompts, knowledge base size, and conversation length. To give a rough sense, here's the approximate cost of a **single reply** that reads \~1,500 tokens of context and produces a short (\~250-token) answer: | Model | Tier | Approx. credits per reply\* | | ------------------- | ----------- | --------------------------- | | `DeepSeek-V4-Flash` | Lightweight | \~0.3 | | `gpt-4o-mini` | Lightweight | \~0.4 | | `gpt-5.4-nano` | Lightweight | \~0.6 | | `gpt-oss-120b` | Lightweight | \~0.7 | | `gpt-5-mini` | Lightweight | \~0.9 | | `gemini-2.5-flash` | Balanced | \~1.1 | | `Qwen3.6 Plus` | Balanced | \~1.5 | | `Kimi K2.5` | Balanced | \~1.7 | | `gpt-5.4-mini` | Balanced | \~2.3 | | `Kimi K2.6` | Balanced | \~2.4 | | `Grok 4.3` | Balanced | \~2.5 | | `gpt-5` | Advanced | \~4.4 | | `gemini-2.5-pro` | Advanced | \~4.4 | | `gpt-4o` | Advanced | \~6.3 | | `gpt-5.4` | Advanced | \~7.5 | \**Approximate estimates.* Real usage varies with your conversation length, knowledge base, Tool calls, Actions, Analyze, and how much the model reasons. ### Which model should I choose? Pick based on what your agent does: * **Handling FAQs, simple lookups, or routing to the right team?** → Start with a **lightweight** model. It's fast, cheap, and accurate enough for straightforward tasks. * **General customer support or sales over your knowledge base?** → Use a **balanced** model. This is the best starting point for most agents. * **Complex, multi-step questions, nuanced judgment, or high-stakes replies?** → Use an **advanced** model, and only where you need it. * **Very high message volume and tight on credits?** → Favour lightweight models and keep answers concise. ### A note on accuracy Accuracy isn't decided by the model alone. For most Gallabox agents, the biggest factors are: * **The quality of your knowledge base** — clear, well-structured, up-to-date content matters more than the model. * **Your instructions / prompt** — specific, unambiguous instructions improve every model. * **Your use case** — a lightweight model can be *more* than accurate enough for simple tasks, while complex reasoning benefits from an advanced model. Because of this, **we recommend testing 2–3 models with your own knowledge base** and comparing the answers before committing. ### Tips to reduce credit usage * **Match the model to the task** — don't use an advanced model for simple FAQs. * **Keep your knowledge base focused** — smaller, relevant context means fewer input tokens per reply. * **Ask for concise answers** in your agent's instructions — shorter replies use fewer output tokens. * **Use lower reasoning effort** where the model supports it, for simpler tasks. * **Monitor your AI Wallet** — review your usage history to see how each agent and model consumes credits, and adjust. ### Frequently asked questions **What is a token?** A token is a small chunk of text — roughly 4 characters, or about ¾ of a word. Models read and write in tokens, and credits are charged per 1,000 tokens (≈ 750 words). Longer prompts, bigger knowledge bases, and longer answers all use more tokens. **Which model uses the fewest credits?** Lightweight models (like the "nano" and "mini" options) use the fewest. They're ideal for simple, high-volume tasks. **Which model is the most accurate?** It depends on your use case and knowledge base. Advanced models handle complex reasoning best, but for everyday support a balanced model is usually accurate enough — and cheaper. **Can I change my agent's model later?** Yes. You can switch models anytime and test which works best for your needs. **Will a more expensive model always give better answers?** No. For simple tasks, a lightweight model can perform just as well at a fraction of the credit cost. Test before assuming you need the most advanced option.