> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tradionlabs.com/llms.txt
> Use this file to discover all available pages before exploring further.
# AI models
> Which model powers which feature, why the agent model differs by plan, and what is sent to model providers.
Tradion does not run one model behind everything. Different jobs go to different models, chosen for what that job needs — depth of reasoning, the ability to read an image, or speed.
A **model** here means the AI system that reads a prompt and writes the response. Tradion calls them over an API; none of them run on your machine.
## Which model runs which feature
| Feature | Model | Why this one |
| --------------------------------------------------------- | ------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- |
| AI Quant Research — generating the Python for a cell | Claude Opus 4.8 | The hardest job on the platform. Code that has to run first time against real data |
| AI Agent node in automations | Claude Sonnet 4.6 on Trader, Claude Opus 4.8 on Quant. Not available on Starter | See below |
| Chart Analyzer | Claude Sonnet 4.6 | Reads the chart image and returns levels and a verdict |
| Lens | Claude Sonnet 4.6 | Same vision capability, applied to a screenshot you capture |
| Earnings Spider | Claude Sonnet 4.6 | Reads long PDFs and image-based press releases |
| Trade Autopsy | Claude Sonnet 4.6 | Pulls structured trade details out of text you paste, then writes the report |
| AI Asset Manager | Claude Sonnet 4.6 | Conversational, with tools for live market data and your positions |
| Tradion Memory — trader profile and psychological profile | Claude Sonnet 4.6 | Long structured output built from many inputs |
| Terminal follow-up chat | Claude Haiku 4.5 | Short answers about work already on screen. Speed matters more than depth |
| Pattern explanations, report summaries, briefings | Claude Haiku 4.5 | Short, high-volume text |
| News sentiment scoring | Gemini Flash | High request volume across many headlines at once |
| Repairing a quant cell that failed twice | Gemini | A different model family gives a genuinely different attempt |
| Search inside your own history | OpenAI `text-embedding-3-large` | Turns text into numbers so past trades and notes can be matched by meaning |
## Why the agent model differs by plan
The **AI Agent node** — the step in an automation that researches a triggered symbol before you are notified — runs on **Claude Sonnet 4.6 on Trader** and **Claude Opus 4.8 on Quant**.
**Starter has no agent model**, because Starter has no automations. There is nothing to configure and nothing to compare.
This is the only model difference between plans. Every other feature in the table above uses the same model on every plan, including on Starter.
The reason is cost. An agent run is expensive, and Opus costs several times what Sonnet costs for the same work. Quant includes 500 agent runs a month against Trader's 50, so the higher price supports the more capable model.
**What you get for it:** Opus reasons further through multi-step questions and holds more context before writing its answer. On a straightforward "RSI crossed 30, tell me why" run, the two produce similar work. The gap widens when the agent has to chase several tools and reconcile what they return.
## Vision models
Chart Analyzer, Lens, and Earnings Spider all send an image to the model rather than text. Sonnet 4.6 reads the image directly — the candles, the axis labels, the drawn lines, the table in a PDF.
Two things follow from that, and both are worth knowing before you blame the analysis:
* **Resolution matters.** A blurry or heavily compressed screenshot gives the model less to read. A crisp capture of a smaller region beats a fuzzy capture of a whole screen.
* **It reads what is on the chart.** If your indicators are hidden, the model cannot factor them in. Tradion adds live market data alongside the image, but the visual read comes from the pixels you send.
## Structured outputs
Most Tradion features do not ask a model for prose. They ask for a fixed shape — a direction, a conviction number, a set of named sections, a list of risks — and the request declares that shape up front so the response has to match it.
That is why verdicts look consistent from one run to the next, and why the automation webhook has a payload you can write code against instead of a paragraph you would have to parse. It also means a model that cannot answer returns an empty or refusing structure rather than an invented one.
## What is and is not sent to model providers
Sent, when the feature needs it:
* The text of your question or the automation's instruction
* The market data Tradion fetched for that request — prices, indicator values, filings, headlines
* Images and documents you upload to Chart Analyzer, Lens, or Earnings Spider
* The trade details you log, when you generate an autopsy
* Your Tradion Memory profile — a behavioural summary derived from your own trades — where personalisation is switched on
Not sent, because Tradion never holds it:
* **Your brokerage login.** Brokerage connections are held by SnapTrade. Tradion receives positions and transactions, never credentials.
* **Your card details.** Payment information is held by Stripe.
Automation notifications sent to shared destinations — Discord, Telegram, and your own webhook — have behavioural personalisation removed by default before delivery, so your scores and patterns do not land in a channel other people read. In-app and email notifications keep it.
Model providers process this data under their API terms. Tradion's own handling is set out in the [privacy policy](https://tradionlabs.com/privacy).
What goes into your trader profile and how it reaches each prompt.
Where the Sonnet-to-Opus switch sits, and what else changes.
What the confidence number does and does not mean.
Where the market data in every prompt comes from.