Transparency enables awareness and action! As transparency is currently lacking from AI providers, we built the "Token & CO2 Dashboard": Plugins collect usage data directly and anonymously from AI coding sessions that is then displayed:
- input / output and cache read/write tokens
- by provider, model and time
- along with an emissions estimation
- and the cost we'd pay if it was all usage based pricing.
The dashboard is now Open Source to enable reduction action in other companies as well.
Project Goals
Develop more sustainable products
Generate data and insights
Generate ideas and encourage knowledge sharing
Improve processes and organisational structures
Raise awareness and shift mindsets within the organisation
Reduce the organisation's environmental footprint
services or business models
Focus Area
Decarbonisation & Reporting
Energy
What gave you the courage to start?
Mainly the support from my management! My manager and me both knew that it's irresponsible to have everyone use as much AI as they wanted without managing the use, both from a financial and ecological perspective.
The Outcome
We have the dashboard up and running in our company. Usage, emissions and cost are now very transparent for everyone in the company. We've already had great discussions amongst engineers around the best ways to reduce emissions and talk about experiences with tools to reduce token consumption etc. in a regular meeting once a month. Me and my management have since created a strategy to automate waste reduction in the long term.
The Impact
The dashboard allows us to make experiments that will generate knowledge around which models are best suited for which kinds of coding tasks. Our plan is to then use that knowledge to create a tool that will automatically route AI tasks to the smallest model that can handle the task.
I also hope that a lot of other companies will use the dashboard to make their AI consumption visible to their employees and work on reduction strategies.
The Biggest Challenge
The carbon estimation is very tricky as you get basically nothing on that side from the providers. We took a study as our base and derived emission factors per AI model from this. Still, there's a lot of uncertainties around this and we've so far only solved this for Anthropic models. The dashboard can show consumption across different providers but people who use other providers than Anthropic will have to derive factors for those other providers themselves.
The Biggest Learning
Definitely the fact that in agentic coding, cached tokens can by far be the biggest factor - 97-99% of totals if done right. That also means that emissions from coding sessions are a lot lower than what we feared at first.
Most Enjoyed
I loved working together with other colleagues that usually don't work on sustainability related projects, and to see how they got enthusiastic and interested in this as well. Also, the public interest after Open Sourcing the project was quite big, compared to what we're used to normally. It was great to see that we'd created something that would provide value to many others as well.