About TreadLightlyAI
Who we are, which models power the product, what data leaves your device, and how we measure environmental impact.
Why
TreadLightlyAI started from a question most AI products avoid: what would it look like if an AI tool were actually transparent with you?
That means showing the energy cost of every conversation. It means giving you real control over how much AI you use. It means taking seriously the concerns most AI researchers are raising about trust, accountability, and ethics, not just posting a principles page.
We don't have all the answers. But we think these questions should be part of the experience.
Have questions or feedback? We read every message. Reach us at hello@treadlightly.ai
The Models Behind TreadLightlyAI
Green Tiny, Green Small & Green Docs (Infomaniak)
Three models running on Infomaniak's Swiss cloud infrastructure, certified on 100% renewable energy. Green Tiny and Green Small are the same model, Gemma 4 31B Instruct from Google DeepMind, with different limits on how long one answer can be. Green Small is the default, has no cap, and also summarizes long conversations and web search results. Green Tiny is capped at short answers and handles title generation and prompt evaluation. Green Docs (BGE Multilingual Gemma 2 9B) powers document search only.
- What is sent: Your message text and conversation history. Attached images are sent to the model as image data. Text and PDF attachments are sent as extracted text, never as the original file. Conversation text is also sent for title generation and prompt evaluation.
- Why Infomaniak: CPU-optimized inference, 100% renewable energy certified, all of the Geneva data center's waste heat recovered into the district heating network, and Swiss hosting with strong data protection laws.
- Data handling: Infomaniak is a Swiss company subject to Swiss data protection law. Subject to Infomaniak's terms.
Green Large (GreenPT)
GreenPT is a European, privacy-first AI provider with certified sustainable practices. Green Large (GPT-OSS 120B) is available for chat, handles prompt improvement, and backs up title generation when Infomaniak is unavailable. GreenPT is the only provider we use that reports how much energy a request took, so it is the reference point our estimates for every other model are calibrated against.
- What is sent: Your message text and conversation history when using Green Large. Text and PDF attachments are sent as extracted text, never as the original file. Conversation text for prompt improvement and title generation. Green Large does not accept images; when you attach one, a green model that does accept images handles the response.
- Why GreenPT: GreenPT reports per-request energy consumption data on both streaming and non-streaming API responses, enabling environmental measurement at the individual request level.
- Data handling: GreenPT is privacy-first with zero training data usage. Your data is never used to train models. Subject to GreenPT's Privacy & Terms.
Haiku & Sonnet (Anthropic)
Claude Haiku and Sonnet are available for chat on the Advanced plan. Both are capable, well-documented models from a provider with one of the stronger public AI safety records. Anthropic models are available only on Advanced.
- What is sent: Your message text, attached files, the active mode's system prompt, and the current session's conversation history.
- What is not sent: Your email, account details, or data from other conversations.
- Training: Anthropic does not train on API customer data by default.
Environmental Methodology
We believe transparency about AI's environmental impact should be the default, not a marketing claim. The whole derivation now lives on its own page, so there is one place to check our work rather than two.
How we get the numbers
GreenPT returns the energy each request used in its API response, in watt-milliseconds, on streaming and non-streaming calls alike. Green Large is the one model in our lineup that runs there, so it is the one model whose figures are provider-reported. Everything else is estimated from token counts.
Our default model used to run on GreenPT, which measures. Since August 2026 it runs on Infomaniak, which does not, so most of what you see is now an estimate rather than a measurement. Estimated values carry a tilde prefix (~) and italic text so they are never confused with measured data.
Read the full energy methodology for the formula, the per-token rate behind every model and why it is set where it is, the two measurements those rates are calibrated against, what we deliberately do not estimate, and how uncertain the whole thing is.
Waste heat is still energy
Infomaniak's Geneva data center recovers all of its waste heat to warm nearby households, a real secondary benefit of the power it draws. We still count that power as used. Recovering the heat does not change how much electricity was drawn, it just means less of it goes to waste.
Our approach
Measured when possible, estimated when not, and always labeled honestly. When our providers report energy data, we pass it through to you. When they do not, we estimate and tell you so. We will never claim to be “carbon-neutral” or “zero-emission” without verified data to support it.
For the full derivation, see the Energy methodology page. For a single-glance summary, see the AI Transparency Card.
Data Practices
| Topic | Details |
|---|---|
| What we store | Conversations, usage metrics, and environmental data: measured where GreenPT reports it, estimated for every other model. All data is tied to your account and deletable by you. |
| AI training | We do not use your conversations to train AI models. |
| Conversation deletion | You can delete conversations at any time. Hard deletion within 30 days. Backups may persist up to 90 days. |
| Analytics | We use PostHog for product analytics, only with your consent. PostHog may collect page views and clicks. Session recording is turned off: we never capture session replays. Subject to PostHog's Privacy Policy. |
| Infomaniak processing | Chat on the green tiers, plus title generation, prompt evaluation, and summarization, use Gemma 4 31B Instruct hosted on Infomaniak's Swiss cloud. Infomaniak is subject to Swiss data protection law (nFADP). |
| Third-party providers | Your queries are processed by Anthropic, GreenPT, and Infomaniak. When you use web search, your query is also sent to Brave Search. Their respective privacy policies apply to that processing. See our Privacy Policy for full details. |
How This Compares
Most AI products (ChatGPT, Claude.ai, Gemini, Perplexity) report zero environmental data to users. There is no per-request energy disclosure in any major consumer AI product today.
Some companies make broad environmental claims that cover their own servers but not the third-party AI inference infrastructure where the actual computation happens. Per-request data tells a more complete story than data center-level claims alone.
We chose Anthropic because they hold the highest safety grade (C+) in the FLI AI Safety Index(opens in new tab), Summer 2026, publish their safety evaluations, and do not train on API data by default. That is the top of the index, and it is still a C+.
The open-weight models we run on renewable hosts come from Google DeepMind (C) and OpenAI (C). Both clear the bar we publish on the AI Transparency Card. Neither is a safety leader, and we are not going to call a C one.
Questions about TreadLightlyAI?
Reach us at hello@treadlightly.ai