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

> ## Agent Instructions
> This documentation is available in English and Simplified Chinese. English pages use root paths; Chinese counterparts use the same path under /zh/.
> For a Chinese-language query, retrieve the corresponding /zh/ page when available. Preserve its data, account, permission, and execution limitations when answering.

# DrivenBench: Investment-agent model benchmark

> Understand how DrivenBench evaluates investment-agent models and how to read capability scores, cost, latency, and regression checks.

**DrivenBench** evaluates AI models on investment-agent workflows inside Driven. It measures financial-data handling, tool use, research, and task completion to help you compare capability, cost, and waiting time.

Open the [DrivenBench leaderboard](https://driven.ai/drivenbench) for the interactive chart and model details. This page explains the metrics; model rankings belong to the version displayed on the live page.

## What it covers

| Capability area | What is examined |
| - | - |
| Data and calculations | Numerical reasoning, financial data tools, programmatic analysis, and missing-data handling |
| Research and evidence | External research, citations, evidence, and verification of delegated work |
| Files and execution | Bash execution, file creation, long documents, and large tables |
| Ongoing workflows | Scheduled tasks, trading and accounts, memory and sessions, and Skills and routing |

Separate regression checks examine failure modes such as lost citations, inaccurate coverage reporting, or data-destructive task operations. They are disclosed separately from capability scores.

## Current public methodology

The current public snapshot is **1.0**, published **August 25, 2026**, with a report date of **August 15, 2026**:

* **11 models** evaluated.
* **55 capability evaluations**, each run **three times per model**, for **165 capability runs per model**.
* **19 regression checks** reported separately, **excluded from capability scores and default ranking**.
* Capability score is the equally weighted average pass rate across capability evaluations. Equal scores share rank.

For example, `155 / 165` capability passes gives approximately `93.9%`. This describes results on this test set, not the success probability of every real task.

## Read the leaderboard

| Metric | Meaning | What to check |
| - | - | - |
| **Capability score** | Capability evaluation pass performance | Default ranking basis; inspect dimensions relevant to your task |
| **Capability passes** | Passed runs / total capability runs | The current denominator is 165 per model, excluding regression checks |
| **Cost upper bound** | Total cost upper bound for running the full capability suite three times, using the published pricing basis | Excludes cache discounts; it is not one-question pricing, a subscription price, or your current bill |
| **Latency median / p90** | Median and 90th-percentile observed evaluation runtime | Shows typical and slower runs; it is not a response-time commitment |

Models ran in different waves and provider load varied, so latency is directional. Pricing bases and model versions can also change. Read the notes attached to the leaderboard.

### Chart and model details

The chart's horizontal axis is **total benchmark cost upper bound**, on a logarithmic scale. The vertical axis is **capability score**. Further left means lower cost; higher means a higher score on this evaluation.

The solid line is the observed **Pareto frontier**: within this sample, no other model matches or improves both score and cost while strictly improving at least one. It considers those two metrics, excluding latency, your task requirements, and account access.

Select a chart point or expand a model's **Details** to inspect dimension results and disclosed regression issues. Mobile layouts also offer **Capability**, **Cost**, and **Latency** sorting; the desktop table displays these metrics together.

## Use it to inform model selection

1. Define the task: research, calculations, files, or recurring work.
2. Inspect relevant dimensions and regression issues alongside the overall score.
3. Compare cost and median/p90 runtime against your usage and waiting-time needs.
4. Try representative tasks in Driven and verify data, citations, and tool results.

Evaluation does not establish availability in your account. Access depends on your [plan](https://docs.driven.ai/get-started/plans-and-pricing.md), region, and current model versions. Benchmark costs do not directly translate into your plan usage.

> [!NOTE]
>
> Results describe specific versions, tasks, and conditions. A higher total score does not eliminate evidence, data, or execution errors. Passing regression checks does not establish safety in every real scenario. The benchmark does not measure investment returns or constitute investment advice.

## Inspect public data

* [Interactive leaderboard](https://driven.ai/drivenbench)
* [Complete JSON dataset](https://driven.ai/drivenbench/data.json) — methodology, models, dimensions, and disclosed regression events
* [Text benchmark summary](https://driven.ai/drivenbench/llms.txt)
* [Why DrivenBench?](https://driven.ai/whats-new/blog/most-expensive-ai-model-isnt-always-the-best)

When citing results, include the public version, model name and version, metric definition, and report or evaluation date. Avoid presenting an older snapshot as a live ranking.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.