It is easy to make the model decision feel like the whole AI strategy.

That is understandable. LLMs are visible, fast-moving, and full of vendor claims. But for a bank building an investing AI agent, the more useful question is not simply which model is best? It is which model strategy lets us stay compliant, accurate, adaptable, and commercially sensible as the market changes?

That is why Charlie is model-flexible. You can choose the model that fits your requirements today, and keep the rest of the investing layer stable when better or cheaper models appear tomorrow.

Start with requirements

There is no single best model for every financial institution. The right setup depends on what the agent needs to do, where data may go, and how the experience should behave.

Cost is an architecture question

A good LLM setup can be much cheaper than people expect. With a strong fast model, a normal useful Charlie conversation can often sit around 2 to 8 cents, depending on length, model choice, tool calls, deployment setup, and volume.

Very small interactions can be far lower. In some cases, a conversation can be around 0.01 cents. That is only possible when the product is built carefully: compact prompts, compact tool outputs, smart routing, and no habit of sending the full investing app into the model.

For example, Google’s Gemini API pricing lists Gemini 3.1 Flash-Lite Preview at low per-million-token rates. That kind of pricing makes low-cent conversations realistic, but only if the system uses the model efficiently.

This is one of the reasons Charlie is built around tools. The LLM should communicate, reason, and decide which capability to use. It should not calculate portfolio performance from memory, invent market facts, or receive unnecessary raw data.

Why not build your own model first?

Building your own model is not a bad idea. It can make sense when the model itself is the strategic product, or when there are very specific sovereignty, research, or performance reasons.

For most investing AI projects, it is just not the best first investment. The real work is usually elsewhere: connecting portfolio data, broker systems, market data, deterministic calculations, guardrails, evaluations, traceability, and the operating model your teams can approve.

Training or deeply owning a model also creates ongoing work: specialist talent, infrastructure, data governance, red teaming, monitoring, model updates, incident response, and proof that the model behaves safely over time. That can be worth it, but it should be a deliberate decision, not the default starting point.

Keep the model layer movable

Model switching should still be tested carefully. Different models behave differently. Tool-calling, tone, latency, refusal style, multilingual quality, and safety behaviour all need review.

But a model change should be a contained project, not a rebuild. In Charlie, the portfolio tools, market-data connectors, guardrails, traces, and workflows stay stable while the model layer can evolve.