FAQ

Does the gateway store my prompts?

By default, no. Fairmeter stores request metadata (timestamps, model and provider, token counts, latency, cost, and the routing decision), which the owning team can browse in the requests browser. Prompt and completion payloads are not logged unless your team explicitly enables payload logging. When it is enabled, payloads are visible only to authorized team members and are kept for your team’s configured retention window, and payload logging can be disabled again at any time.

Does Fairmeter offer Zero Data Retention (ZDR)?

Yes. ZDR is an org-level setting you turn on in Settings, and with it on, request and response bodies are never persisted and payload capture cannot be armed while it is on. Metadata rows (timing, token counts, cost, model chosen) are still written, because billing and the usage page run on them, so this is not a claim that nothing is stored. See Security for the exact boundary, including how conversation state and response caching behave under ZDR.

Do you train on my prompts?

Not through any automated pipeline. Fairmeter keeps an internal, manually-run export script an engineer can use to build a training set for the routing classifier, and it only draws from payloads your team has already chosen to retain: if payload logging is off (the default) or Zero Data Retention is on, nothing is written to storage for it to read. When payload logging is on, retained rows carry a sensitivity tag that defaults to internal, which the export treats as eligible, and there is currently no per-request control to mark a row confidential instead. The script is not scheduled and does not run automatically; keeping payload logging off, or turning on Zero Data Retention, keeps your prompts out of it entirely.

How do I tell which rate limit a 429 hit?

Every 429 response carries an X-RateLimit-Scope header naming the limit that fired. The scope values are rps, rpm, tpm, daily_budget, monthly_budget, and key_spend_ceiling, so a client can back off on a burst and escalate on a blown monthly budget without opening a support ticket. Retry-After and X-RateLimit-Reset ride along on the rps, rpm and tpm windows. Read the value rather than assuming it: rps is a one-second per-organisation burst ceiling and says to wait a second, while rpm and tpm run on the sixty-second window. A spend cap has no rolling reset, so do not assume those two headers are always present.

Which models are supported?

Three providers: OpenAI, Anthropic, and Google. On OpenAI that is the GPT-5 line, including the reasoning models, with GPT-4o and the o-series kept as deeper fallback rungs. On Anthropic it is the Claude family (Opus, Sonnet, Haiku). On Google it is Gemini Flash. The routing ladder tracks each provider’s current frontier model, so a roster refresh does not change your call site. A separate model, claude-fable-5, is reachable only through the cap:orchestrate capability handle rather than through the ladder above; see Capability aliases.

Can I bring my own provider keys?

Yes, and you should. Per-org provider credentials live in /providers, encrypted at rest. The gateway uses your keys for upstream calls and bills you against the provider’s actual usage, so the gateway stays revenue-neutral on the upstream side.

Is there a free tier?

No, but the entry point is low: Basic is $9.99/mo, self-serve, on BYOK. Fairmeter is not open source and there is no self-hosted distribution today. A $70/mo Team plan adds per-team budgets, the audit log, alerting, org-scoped access controls, and custom band maps. Enterprise covers on-prem or VPC deployment and a SOC2 evidence pack on request. See Pricing.

How do I correlate a gateway error with a request?

Every response includes an X-Request-Id header. Open /requests/<id> to see the full pipeline trace.