# Evaluating Pareto on Hack The Box Challenges

> Source: <https://theaq.blog/2026/09/20/evaluating-unbiased-pareto-on-hack-the-box-challenges.html>
> Published: 2026-09-20 11:02:41+00:00

# Evaluating Pareto on Hack The Box Challenges

Have you ever heard of [Unbiased AI](https://unbiased.ai/) or the **Pareto** model that they are working on? Me neither. That’s why I was quite surprised when this model performed very well in my testing. And I was even more surprised when I found out that Unbiased AI is not an LLM-making company and that **Pareto** is not actually an LLM.

This blog post is part of a series of tests for the [HTB-Challenger Benchmark](https://theaq.blog/htb-challenger-benchmark/)The HTB-Challenger Benchmark evaluates LLMs’ ability to find and exploit security vulnerabilities. It tests models against selected Hack The Box challenges of varying difficulty and measures their performance. For more information, visit the [HTB-Challenger Benchmark page](https://theaq.blog/htb-challenger-benchmark/).
.
See the [benchmark results page](https://theaq.blog/htb-challenger-benchmark) for all results and the [benchmark methodology](https://theaq.blog/htb-challenger-benchmark-methodology) to learn how the benchmark is calculated.

Let’s clear things up a bit: **Pareto** is a “blended AI model”. You can read about how it works in [this article](https://unbiased.ai/how/), but basically **Pareto** is just a facade for a platform consisting of multiple (undisclosed) LLMs. You send it a message, it is processed by a few LLMs, their answers are aggregated, and the result is sent back to you. This is not a new concept; for example, OpenRouter introduced a similar [Fusion](https://openrouter.ai/openrouter/fusion) “model” some time ago, but I haven’t tested it yet.

My verdict based on this testing is this: in terms of performance, **Pareto** is excellent at cracking Hack The Box challenges and achieved a score of 80.8% in my benchmark. It’s not as good as [GPT-5.6 Sol](https://theaq.blog/2026/08/16/solving-htb-challenges-with-openai-gpt-5.6-sol.html), which scored 87.2%, but it’s on the same level as [GLM 5.3 Flash](https://theaq.blog/2026/08/28/evaluating-z-ai-glm-5.3-flash-on-hack-the-box-challenges.html) (81.0%) and [DeepSeek V4.1 Flash](https://theaq.blog/2026/09/11/evaluating-deepseek-deepseek-v4.1-flash-on-hack-the-box-challenges.html) (80.9%).

Its only issue is that it’s relatively expensive. Its median cost per challenge was $0.37, which is not only MUCH higher than **GLM 5.3 Flash** ($0.01) and **DeepSeek V4.1 Flash** ($0.05), but also higher than **GPT-5.6 Sol** ($0.29).

So, good job on the performance, Unbiased AI, but now you need to fire a few managers and janitors to make your company more cost-effective.

## Cost vs. Benchmark Score

The highlighted point is this model. Models closer to the upper-left achieve a higher benchmark score at a lower median cost per challenge.

## Overall benchmark results

- **Number of challenges:** 16
- **Number of solved challenges:** 15
- **Number of false positives:** 0
- **Runs where the model gave up:** 1
- **Runs that reached the step or cost limit:** 0
- **Runs where the model got stuck:** 0
- **Benchmark score:** 80.8%

| Metric | Per challenge (median) | Total | 
|---|---|---|
| Model steps | 25.5 | 555 | 
| Model cost | $0.37 | $21.43 | 
| Duration | 00:08:22 | 03:32:18 | 
| Number of input tokens | 0.34M | 17.50M | 
| Number of output tokens | 0.02M | 1.70M | 
| Number of `read_file` tool calls | 1.0 | 31 | 
| Number of `write_file` tool calls | 4.0 | 90 | 
| Number of `execute_command` tool calls | 23.5 | 592 | 
| Number of `web_search` tool calls | 0.0 | 1 | 

## Results by challenge difficulty

All resource-usage metrics are medians per challenge.

| Metric | Very Easy | Easy | Medium | Hard | 
|---|---|---|---|---|
| **Results** |  |  |  |  | 
| Number of challenges | 4 | 4 | 4 | 4 | 
| Number of solved challenges | 4 | 4 | 4 | 3 | 
| Number of false positives | 0 | 0 | 0 | 0 | 
| Runs where the model gave up | 0 | 0 | 0 | 1 | 
| Runs that reached the step or cost limit | 0 | 0 | 0 | 0 | 
| Runs where the model got stuck | 0 | 0 | 0 | 0 | 
| Benchmark score | 95.9% | 92.0% | 91.2% | 63.6% | 
| **Median per challenge** |  |  |  |  | 
| Model steps | 14.5 | 23 | 29 | 52 | 
| Model cost | $0.20 | $0.26 | $1.04 | $2.52 | 
| Duration | 00:01:57 | 00:06:18 | 00:11:29 | 00:22:28 | 
| Number of input tokens | 0.14M | 0.19M | 1.12M | 1.79M | 
| Number of output tokens | 0.01M | 0.02M | 0.07M | 0.19M | 
| Number of `read_file` tool calls | 0.5 | 1.0 | 0.5 | 1.5 | 
| Number of `write_file` tool calls | 1.0 | 6.0 | 4.0 | 7.5 | 
| Number of `execute_command` tool calls | 12.5 | 17.0 | 28.0 | 51.0 | 
| Number of `web_search` tool calls | 0.0 | 0.0 | 0.0 | 0.0 |