No more per-vendor sign-ups and key sprawl — one key reaches 9 models. Wire-compatible with OpenAI, so existing SDKs keep working untouched; unified USD settlement at 20% off of list price, with automatic upstream failover.
No new SDK, no code changes, no per-vendor keys.
Email or social login; one key unlocks every vendor’s models — no card required.
sk-····a41f
Keep your SDK and code; change exactly two environment variables.
OPENAI_BASE_URL= https://compassapi.ai/v1
Swap the model name for a unified slug and switch models anytime.
model="gpt-5.4-mini"
How fallbacks, caching and settlement work — at a glance.
Every vendor behind one base_url and one key, with unified slugs — no per-vendor keys to request, no re-architecting to switch models.
Each model rides multiple upstream channels, weighted; timeouts and rate limits retry the next channel with no caller impact.
Repeated input prefixes bill at 20%; retrieval-augmented apps typically save a third of their cost.
One pricing basis for every model, settled in USD — no juggling exchange rates and tier tables across vendors.
Only model, tokens, latency and request id are recorded — never request or response bodies. Compliance reviews pass easily.
Every key gets its own spend cap and model allowlist; a leak or a runaway loop cannot burn through the account.
Upstream channels, routing policy and billing logs live in one console — no more per-vendor detective work.
Same list price as going direct — the savings come from cache-hit discounts and retries you no longer pay for. Drag to see your monthly bill.
Estimated from public unit prices and the platform-average cache discount; your usage log is authoritative.
USD per million tokens; cache-hit input bills at 20%.
Any client or framework with a custom OpenAI base_url just needs the address and a key.
Fields not listed pass through to the upstream unmodified.
No. Upstream channels are maintained by the platform — one key and one base_url from this site reach every connected vendor’s models, and newly added vendors work with nothing but a model-name change.
Each model carries its own discounted price. The struck-through figure in the price table is the upstream list price and the one beside it is what you pay — no account tiers. The rest of the savings come from cache-hit discounts and retries you no longer pay for.
Prefix caching is detected automatically; the cached share of input bills at 20% of list, cache writes at 1.25×. cached_tokens shows as its own line in the usage log, with the full formula in request details.
Two environment variables: point base_url at this site and swap in your key. Streaming, function calling and vision inputs are wire-identical to OpenAI.
Each model rides multiple upstream channels; timeouts and rate limits retry the next channel with no caller impact, and edge-node failures fail over to same-region backups.
Only model, token counts, latency and request id — never request or response bodies. Compliance reviews pass easily.
No. Sign in with email or a social account and start calling; top up only when you need more quota.