What an AI API Relay Is and How to Choose One
Understand how AI API relays differ from official APIs, then compare exact models, protocols, route classes, input and output rates, paid requests, and billing evidence.
Definition
An API relay adds an intermediary service between a client and one or more model upstreams. It may unify base URLs, keys, protocols, routing, and billing, but it is not the model vendor's official API.
Why the lowest rate is not enough
The same model name may appear on direct, pooled, reverse, or client-specific routes. Input, output, cache, and subscription quotas also use different billing rules.
What we verify
Our six paid test accounts retain model-level successes and failures, end-to-end response time, provider-owned price sources, and unresolved checks. Website uptime, API success, model identity, and billing remain separate evidence layers.
Official API and relay API are not the same product
| Item | Official API | AI API relay |
|---|---|---|
| Operator | The model vendor | An independent relay operator plus one or more upstreams |
| Endpoint and protocol | Vendor-native endpoint and protocol | May be vendor-native, OpenAI-compatible, translated, or client-specific |
| Pricing | Vendor list price and account terms | Provider rates, route multipliers, subscriptions, credit, and cache rules |
| Evidence needed | Official documentation and account bill | Provider source plus model-level calls, limits, and balance reconciliation |
| Additional risk | Vendor availability and account policy | Relay availability, upstream changes, support, privacy, and prepaid balance |
A relay can reduce integration or payment friction, but compatibility, price, upstream identity, and long-term availability are separate claims. None should be inferred from a homepage or model-list response.
A five-step relay selection workflow
1. Name the exact workload
Write down the model ID, API protocol, client, streaming or tool requirements, and expected monthly input and output tokens.
2. Keep route classes separate
Compare provider-labeled official or direct routes with the same class. Keep pooled, reverse, Kiro, Claude Code, and Codex-only routes in separate groups.
3. Verify generation, not only model discovery
A model-list response can pass while generation fails. Run the exact protocol and model you plan to use, and retain both successful and failed samples.
4. Reconcile the wallet
Record balance before and after a fixed request together with input, output, and cache tokens. Published rates and full-use subscription equivalents are not observed charges.
5. Check failure handling
Confirm support, refund rules, daily quotas, concurrency, expiry, and what happens when an upstream model or route changes.
Same-model price and paid API evidence
Six provider profiles: compare input and output rates for the same model, then inspect the actual request outcomes. Live rankings and subscription estimates may differ from this build-time snapshot.
Claude Opus 5 · premium routes
| Provider | Price group | Input / 1M | Output / 1M | Source date | API passed / total | Latest test | P50 total | Best suited for |
|---|---|---|---|---|---|---|---|---|
| LinksAPI | Dedicated CC Max | ¥3.4 | ¥17 | 2026-09-25 | 4 / 4 | 2026-09-29 | 3.48 s | Users prioritizing a provider-labeled premium route |
| H API | Claude Max | ¥6.4 | ¥32 | 2026-09-25 | 3 / 3 | 2026-09-25 | 6.72 s | Claude Code or VS Code users who accept a client-only route |
| UU API | CC full-capability MAX - CC client only | ¥6.8 | ¥34 | 2026-09-25 | 1 / 3 | 2026-09-25 | 3.61 s | Claude Code or VS Code users who accept a client-only route |
| APINebula | Claude Code global protected route | ¥6.8 | ¥34 | 2026-09-25 | 3 / 4 | 2026-09-29 | 2.73 s | Claude Code or VS Code users who accept a client-only route |
| boxying | Claude official route | ¥8 | ¥40 | 2026-09-25 | 2 / 11 | 2026-09-25 | 2.66 s | Users prioritizing a provider-labeled premium route |
| OpenOx | direct | ¥19.2 | ¥96 | 2026-09-24 | 3 / 3 | 2026-09-27 | 3.81 s | Users prioritizing a provider-labeled premium route |
Prices are provider-published pay-as-you-go rates, not subscription equivalents or reconciled charges. API samples cover the named model, but the test key's actual upstream group is not consistently recorded and may differ from the quoted group. P50 is successful end-to-end response time, not time to first token. A completed call does not prove model identity.
GPT-5.6 Sol · premium routes
| Provider | Price group | Input / 1M | Output / 1M | Source date | API passed / total | Latest test | P50 total | Best suited for |
|---|---|---|---|---|---|---|---|---|
| H API | GPT Pro | ¥1.5 | ¥9 | 2026-09-25 | 3 / 3 | 2026-09-27 | 3.14 s | Codex or Responses API workloads after protocol testing |
| APINebula | CODEX | ¥1.95 | ¥11.7 | 2026-09-25 | No sample | - | - | Codex or Responses API workloads after protocol testing |
| UU API | Codex-GPT Pro account pool | ¥2 | ¥12 | 2026-09-25 | 3 / 3 | 2026-09-25 | 1.83 s | Codex or Responses API workloads after protocol testing |
| boxying | Codex premium pool | ¥2.2 | ¥13.2 | 2026-09-25 | 3 / 3 | 2026-09-25 | 10.8 s | Codex or Responses API workloads after protocol testing |
| LinksAPI | GPT Pro (Codex included), 0.3x | ¥3.25 | ¥26 | 2026-09-25 | 3 / 3 | 2026-09-27 | 17.0 s | Codex or Responses API workloads after protocol testing |
| OpenOx | Pro direct route | ¥9 | ¥54 | 2026-09-24 | 4 / 4 | 2026-09-29 | 2.73 s | Codex or Responses API workloads after protocol testing |
Prices are provider-published pay-as-you-go rates, not subscription equivalents or reconciled charges. API samples cover the named model, but the test key's actual upstream group is not consistently recorded and may differ from the quoted group. P50 is successful end-to-end response time, not time to first token. A completed call does not prove model identity.
Frequently asked questions
What is an AI API relay?
It is an intermediary endpoint between a client and one or more model providers. It may provide unified authentication, protocol conversion, route selection, and local billing, but its upstream source and restrictions must be checked separately.
Is an AI API relay the same as an official API?
No. An official API is operated by the model vendor. A relay adds another operator, billing layer, and route between the client and upstream service, even when it exposes a compatible protocol or claims an official route.
How should I choose an AI API relay?
Fix the exact model, protocol, client, and route class first. Then compare separate input and output rates, paid request samples, client limits, support terms, and billing evidence. Start with a low-balance key.
Does a successful request prove the model and billing are genuine?
No. It proves only that the tested key and route returned a response at that time. Model identity and the charged amount require separate capability probes and before-and-after balance reconciliation.