Announcing Infracost AI: model your AI spend decisions before you make them

Today, I’m super excited to launch Infracost AI alongside Infracost Cloud. We are starting by helping large enterprises build what-if scenarios to build business cases for switching to open weight models. But first, let me tell you how we got here:

We launched Infracost in 2021 with a simple pitch: the people who are purchasing and using cloud resources are never shown the cost impact of their infrastructure before they purchase them. That is unfair, and we needed to make cloud costs visible to developers.

In order to achieve this, we created integrations with Integrated Development Environments (IDE) like VSCode and JetBrains. We also integrated directly into GitHub, GitLab and Azure DevOps. We then shifted further left and connected directly to coding agents like Claude and Cursor.

Over the last 5 years, we have become the leaders in the whole Shift Left movement in the FinOps landscape. Infracost is now run over 2,000,000 times every month.

We have gained a very unique perspective into AI usage. We sit in the workflow where AI is being used to generate code, and send pull requests. Here is our learning so far:

Tokenomics has become a board-level topic

Code generation tools are being used by engineers at a very high adoption rate. I think we can all safely say we feel this part. However, there are two parts that might be a little less obvious:

1. Every employee in the company is now using AI tools to help them accelerate their work, and create better outcomes.

2. As this rate increases, costs have become a real barrier, companies of all sizes are looking at alternatives to the frontier models.

Here is the usage of different model providers in the first half of 2026 alone:

You can see the massive growth. But you can also start to see DeepSeek, an open-weight model provider, starting to get a big part of the pie. You can also clearly see that the world is going to be multi-model, which is different from cloud days, when 80% of workloads would be on a single cloud provider.

Here is what that looks like in the last month (August 2026), with DeepSeek Flash taking the top of the leaderboard:

Switching out frontier models for open-weight models are predominantly driven by a focus on cost reduction, especially in larger enterprises. Here are two public examples:

• AT&T: 56% savings by swapping Claude for open models

• Pinterest makes 90%+ cost savings by dropping frontier models

This poses a big question for our users:

What is the business case for switching to open weight models?

What would happen to our costs if we were to use more open-weight or open-source models across the board? What would happen if we implemented model routing strategies, where some requests route to the frontier models, and others route to open models running on our own infrastructure?

Introducing Infracost AI:

Infracost Cloud has Shifted FinOps Left, we now need help on the design side of the AI model usage.

Today, I’m super excited to announce the launch of Infracost AI. We are launching it with two main capabilities:

See all usage (both billed and token based) across the organization

First, we need to bring in your historic usage so you can see where your token usage is today. 

There are two types of AI usage: internal and product (external). Internal usage is AI that your teams use to do their work - think coding agents like Claude Code or Microsoft Copilot. Product usage is AI that comes from your product and is used by your customers, such as a chat app on your website powered by AWS Bedrock or LiteLLM.

The challenge is that AI usage data is fragmented. Provider billing formats differ and change constantly, and the same request often shows up in more than one place. Most enterprises we speak with route LLM requests through an AI gateway, so granular usage data lives in the gateway while the invoice comes from the provider. Mapping those two together, and removing the overlap, is hard, and most FinOps and AIOps teams are still scrambling to establish a baseline for org-wide AI consumption.

Infracost AI ingests LLM provider billing and subscription data, AI gateway data, and OpenTelemetry data from coding agents and applications. It then reconciles this information so you can understand what’s being used across the organization - providers, models, people, teams, products, customers, and more. Getting this step right is critical, because every scenario will be built on top of this data:

What-if analysis with Scenario Builder

Now that we have usage data, we can build scenarios. These scenarios will help you:

1. Build a business case for whether using open-weight models would be economical.

2. Build a business case for investing in model routing strategies - for example, routing a specific team’s usage to a specific model or provider while others are not.

3. Create a financial model of what would happen with your current usage, as well as how costs might change in a few months or next year as usage increases.

In order to build this feature, we first had to bring in the price of all models. We are making that free and available to everyone: https://cost.dev/models 

This data is then combined with usage data to create future scenarios.

Our Infracost Cloud product is focused on preventing and remediating cloud costs and policy issues across cloud and code.

And with our new Infracost AI product launch, we give you insights and the ability to create what-if scenarios for AI consumption and usage.

I’m super excited to launch Infracost AI into early access. If you’d like to be a design partner as we build the feature set out, please fill-out this form, and as a thank you, I will personally give you preferential pricing when the product goes to general availability.

Let’s build.

Be a design partner and gain early access: https://www.infracost.io/product/ai