Choosing an AI Model

NinjaTrader Xen offers several AI models through one workspace. The selected model is used for the next request and is remembered when you return. The strongest model is not automatically the best choice for every task; cost, speed, coding ability and image support all matter.

The available list currently includes ChatGPT Sol 5.6, ChatGPT Codex 5.3, Claude Opus 5, Claude Sonnet 4.6, DeepSeek Pro V4 and Kimi K2.7 Code. DeepSeek and Kimi are marked as lower-cost choices in the interface.

Choosing for the work

ChatGPT Codex is a strong choice for compiler errors, targeted code repair and modification of an existing file. ChatGPT Sol and the Claude models are appropriate for larger or more demanding builds. Claude models also support the visual-reference workflows used by Indicators.

DeepSeek Pro and Kimi K2.7 Code provide lower-cost coding options. They are useful for ordinary builds, explanations and revisions, although provider response time and output quality can vary with the complexity of the request.

For Strategy Analyzer analysis, use a capable reasoning model that can compare several performance metrics without overemphasising net profit. The task-specific prompt remains the same whichever supported model is selected.

Model restrictions

Some workflows deliberately keep a proven routing choice. A repair escalation may offer Retry repair with Codex 5.3 after repeated failures rather than silently changing the user’s selection.

Reference-image controls are disabled for models without image input. The interface explains the restriction before the request is sent, so the image is not accidentally discarded.

Credit use

Models charge different amounts for input and output tokens. Large source files, long conversations and extensive generated code consume more credit than concise requests. Lower-cost models are marked in the selector, but exact charges are based on actual usage rather than a fixed fee per message.

Start with a model appropriate to the task and test the result. Switching models is most useful when a particular provider is slow, a complex build is incomplete, or repeated repair attempts have not resolved the problem.