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

# Pricing & Performance

> Understand how model pricing and performance are measured, compared, and normalised across providers on AI Stats.

AI Stats helps you make informed decisions about which models to use by providing **transparent pricing and performance metrics** for every model and provider.

This page explains how AI Stats calculates, standardises, and presents cost and speed data — allowing you to identify the **best value** models for your use case.

***

## Why pricing & performance matter

When choosing a model, you’re balancing **three key factors**:

1. 💰 **Price** — how much each request costs.
2. ⚙️ **Performance** — how fast and efficient the model is.
3. 🧠 **Quality** — how well it performs on benchmarks.

AI Stats gives you clear, comparable metrics for the first two, and links them with benchmark data to help you find the **sweet spot between cost, speed, and intelligence**.

***

## Pricing structure

Pricing varies depending on the **provider**, **model type**, and **usage mode**.\
AI Stats normalises all prices to **USD per 1,000 tokens** (for text models) or the equivalent unit for other modalities.

| Type             | Unit                | Example                              |
| ---------------- | ------------------- | ------------------------------------ |
| **Text Models**  | Per 1,000 tokens    | $0.005 input / $0.015 output         |
| **Image Models** | Per image           | \$0.02 per 1024×1024 image           |
| **Audio Models** | Per minute or MB    | \$0.006 per minute of transcription  |
| **Video Models** | Per second or frame | \$0.01 per second of generated video |
| **Embeddings**   | Per 1,000 tokens    | \$0.0001 per 1K tokens               |

***

## Input vs Output pricing

For most text-based models, you’ll see two separate costs:

* **Input price** — charged per 1,000 tokens you send (the prompt).
* **Output price** — charged per 1,000 tokens the model generates.

Example:

| Model             | Input (USD / 1K tokens) | Output (USD / 1K tokens) |
| ----------------- | ----------------------- | ------------------------ |
| GPT-4o            | \$0.005                 | \$0.015                  |
| Claude 3.5 Sonnet | \$0.003                 | \$0.015                  |
| Gemini 1.5 Pro    | \$0.00125               | \$0.005                  |
| Mistral Large     | \$0.002                 | \$0.006                  |

AI Stats displays both prices for full transparency and allows filtering or sorting by either.

***

## Cost per request (E2E)

AI Stats also computes an **estimated cost per complete request**, which factors in:

* The **average number of input tokens** per request.
* The **average model output length**.
* The **current pricing structure** for that provider.

This gives a more realistic “per-call” cost when comparing across models and providers.

***

## Measuring performance

AI Stats tracks three main performance metrics for every model and provider:

| Metric                 | Description                                                                                        | Example             |
| ---------------------- | -------------------------------------------------------------------------------------------------- | ------------------- |
| **Throughput**         | The average number of tokens processed per second (TPS).                                           | e.g. 400 tokens/sec |
| **Latency**            | The time between sending a request and receiving the **first token** (TTFT - time-to-first-token). | e.g. 800ms          |
| **End-to-End Latency** | The total duration from request to the last token in the response.                                 | e.g. 3.2s           |

Together, these metrics reflect **responsiveness and efficiency** — critical for interactive or production workloads.

***

## How we measure it

AI Stats uses **standardised benchmarking scripts** to gather real-world performance data for each model and provider.

Each test includes:

* Repeated requests with consistent prompts (to reduce variance).
* Averaged results across multiple time periods.
* Measurements from different regions when applicable.

For transparency, raw performance data is often available through the open [Performance Dataset](https://github.com/AI-Stats/AI-Stats) and the catalog endpoints documented in the AI Stats Gateway API reference.

***

## Interpreting performance

| Scenario                         | What to look for                                                                    |
| -------------------------------- | ----------------------------------------------------------------------------------- |
| **Interactive chatbots**         | Prioritise **low latency** (\< 1s TTFT).                                            |
| **Large document summarisation** | Focus on **high throughput** (tokens/sec).                                          |
| **Batch processing**             | Balance throughput and cost — consider cheaper models with slightly higher latency. |
| **Real-time streaming**          | Requires providers that support streaming APIs and sub-second TTFT.                 |

Use AI Stats’ performance charts to visualise these trade-offs interactively.

***

## Example performance snapshot

```json theme={null}
{
	"model": "gpt-4o",
	"provider": "OpenAI",
	"input_price_usd_per_1k": 0.005,
	"output_price_usd_per_1k": 0.015,
	"throughput_tps": 420,
	"latency_first_token_ms": 700,
	"latency_total_ms": 3100,
	"uptime_percent": 99.97
}
```

***

## Cost-to-performance ratio

AI Stats calculates a **cost-to-performance ratio (CPR)** — a derived metric that helps identify the most efficient models.

Formula:

math
CPR = (Cost per 1K tokens) / (Throughput in tokens/sec)

Lower CPR values indicate **better cost efficiency** for the same throughput.

***

## Data refresh frequency

Pricing and performance data are refreshed regularly:

| Data Type       | Update Frequency          |
| --------------- | ------------------------- |
| **Pricing**     | Every 24 hours            |
| **Performance** | Every 6 hours             |
| **Benchmarks**  | Weekly or on new releases |

Each update is timestamped and versioned for transparency.

***

## Example use cases

| Goal                               | Example                                                              |
| ---------------------------------- | -------------------------------------------------------------------- |
| Compare models by speed            | “Which model has the lowest latency?”                                |
| Find the best cost-per-token ratio | “Which model offers the best throughput for under \$0.01/1K tokens?” |
| Identify regional variations       | “Does latency differ between EU and US endpoints?”                   |
| Balance cost and performance       | “What’s the most efficient model for real-time summarisation?”       |

***

## Contributing or validating data

You can help by submitting verified pricing or performance updates through GitHub.\
All submissions are reviewed before being included in the live dataset.

<Card title="Contribute Pricing & Performance Data" icon="github" href="../contributing/overview.mdx" horizontal>
  Help maintain the accuracy of cost and performance data across providers.
</Card>

***

## Next steps

Now that you understand how pricing and performance are measured, you can explore how to integrate models programmatically via the AI Stats Gateway.

<Card title="Integrate with the Gateway" icon="terminal" href="../developers/integrating-with-the-gateway.mdx" horizontal>
  Learn how to build using the AI Stats Gateway API.
</Card>