Estimate from your real usage.

Platform usage vs. consumer apps

Every vendor we support runs two distinct product surfaces: a Platform / Developer API (the thing you call with sk-ant-*, sk-*, or a GCP service-account JSON) and their consumer apps (Claude.ai, Claude Code, ChatGPT, the Codex CLI/IDE, the Gemini app, Cursor Pro / Team seats). Today only the Platform API surfaces expose machine-readable usage; the consumer apps don't.

That means a connector here only covers traffic that went through the Platform API. Specifically, the following surfaces are not covered yet. Token Offset has no way to measure them until each vendor ships a usage endpoint for them:

Anthropic exposes account-wide Claude API usage through its Admin API. To read it we need an Admin API key, a key that starts with sk-ant-admin… and can only be minted by an organization admin. (Standard sk-ant-api03… keys won't work, they can't read org-wide usage.)

1. 1Visit console.anthropic.com → Settings → Admin Keys.
2. 2Click Create Key. Name it something memorable like tokenoffset-readonly.
3. 3Copy the key: it's shown exactly once. If you lose it you'll have to create a new one.
4. 4In Token Offset, open Settings → Estimate sources, pick Anthropic, and paste the key. We validate it against the Admin API on submit; if anything is wrong, you'll see the error inline.

What we read (and don't)

We call exactly three Admin API endpoints:

• GET /v1/organizations/me , to verify the key and pick up your org name for the connection label.
• GET /v1/organizations/usage_report/messages , daily token counts (input, output, cache reads, cache creation) grouped by model, service tier, and context window. We sample 90 days and project to a monthly estimate.
• GET /v1/organizations/cost_report , daily provider spend grouped by description (model, tier, token type), so the projected estimate is grounded in what you actually paid Anthropic over the window.

That's it. We never call the Messages API; we don't see your prompts, your completions, or the contents of any conversation. Anthropic's Admin API doesn't expose them to anyone, including us.

Refresh cadence

On first connect we pull 90 days of history in the background , usually done within a minute or two , and project it to a monthly estimate. After that the sample refreshes quietly once per week on Sundays at 08:00 UTC so the projected number on your dashboard reflects how your usage is trending. You can re-sample anytime from Settings → Estimate sources. The dollar amountwe charge for your offset only updates when you explicitly adopt the new estimate, your billing doesn't drift behind your back.

What we save and what we do with it

For every connection we keep the daily token counts and cost figures we sample, the running 90-day projection, and the timestamp of each sample. We also keep the choices you make on the offset itself, your tier band (Casual / Worker / Agentic / Extreme), any custom amount you set, and whether you opted into the sampled estimate. None of this includes prompts, completions, or conversation content; the vendor admin APIs simply don't expose them.

We use that data three ways. First, to render your dashboard and bill the right amount. Second, to keep our published environmental coefficients and tier midpoints honest, the more real-usage samples we have, the better we can calibrate the bands shown on /docs/methodology. Third, we may share aggregated, de-identified insights (e.g. "agentic-band median tokens grew 18% quarter over quarter") with research, sustainability, and partner initiatives that Token Offset, in its sole discretion, determines to be aligned with the platform's mission. Identifiable data isn't sold; the full specifics live in our Privacy Policy.

OpenAI exposes per-organization Platform API usage through their Admin Usage endpoints. To read it we need an Admin API key, a key that starts with sk-admin-… and can only be minted by an organization owner. (Standard project-scoped sk-… keys won't work, they can't read org-wide usage.)

1. 1Visit platform.openai.com → Settings → Admin Keys.
2. 2Click Create new admin key. Name it something memorable like tokenoffset-readonly.
3. 3Copy the key: it's shown exactly once. If you lose it you'll have to create a new one.
4. 4In Token Offset, open Settings → Estimate sources, pick OpenAI, and paste the key. We validate it against the Admin Usage endpoint on submit; if anything is wrong, you'll see the error inline.

What we read

We call exactly two Admin Usage endpoints:

• GET /v1/organization/usage/completions , daily input / cached / output token counts grouped by model, service tier, and batch flag.
• GET /v1/organization/costs , daily provider spend grouped by line item. We sample it alongside the token counts so the projected monthly estimate is grounded in what you actually paid OpenAI over the window.

Google's Gemini API and Vertex AI both expose usage and cost only through Google Cloud Monitoring and the Cloud Billing API, both of which require a GCP service-account JSON rather than the paste-an-API-key flow we use for Anthropic and OpenAI. There is no per-personal-account "Gemini admin key" that returns daily token counts; AI Studio's personal usage dashboard is HTML-only and not exposed via any public API.

Cursor's agentic and Composer workflows are the largest single source of per-developer token usage on the market, see the methodology page for why. Cursor exposes per-team spending through the Team Admin API ( POST /teams/spend, /teams/daily-usage-data), but the API is gated to Enterprise plans only.