Buyer guide · 2026-11-12 · 321AI engineering notes · get a key
Quick answer: Topping up a relay platform is a trust decision. This checklist gives you nine concrete checks — each one verifiable from public data in minutes — so you can judge a platform on evidence instead of marketing, before a single dollar of balance goes in.
1. Prices you can re-derive. The price table should be backed by a free, open API you can call yourself — no login, CORS-enabled, with an updated_at field. On 321AI that is GET https://321ai.cc/api/public/model-prices. If you cannot reproduce the page numbers from the API, the numbers are decoration.
2. Failure billing, in writing. Somewhere in the product or terms it must say what happens when a call fails. On 321AI: failed requests are never billed and pre-deducted quota is refunded automatically. A platform that is vague about this is planning to keep your money on its bad days.
3. Platform fee arithmetic. Ask what is actually added on top of model cost. Listing fees, commissions and routing fees should each be named. The honest answer on 321AI is a single number: 0.
4. Supplier depth per model. One source means one point of failure. Check the supplier count column on the catalog — models quoted by many lanes survive lane-specific outages without your involvement.
5. Verified supply, backed by something. A "genuine models" promise is only credible with collateral. On 321AI, verified suppliers post a upstream self-check records and mismatches go through a merchant-level support. Collateral turns a slogan into a procedure.
6. Quality metrics in the open. Success rate and P50 latency should be published per model and per supplier, not summarized in a testimonial. You want to pick lanes from data, especially if your app is user-facing.
7. Protocol coverage that fits your tools. If you run Claude Code, Anthropic Messages must be native — not an emulation. Check for both OpenAI protocol (base_url with /v1) and Anthropic protocol (ANTHROPIC_BASE_URL without /v1).
8. Payment and refund path. Top-up channels should be named (on 321AI: USDT across TRON/BSC/Polygon/Solana, plus regional channels shown in the console), and the terms should state that unconsumed balance is refundable. Prepaid should never mean trapped.
9. Support and incident transparency. A ticket system with a stated first-response target (~2-4h on 321AI), a public status page, and incident history. When something breaks — and something always breaks — this is what you will be using.
| # | Check | Evidence to demand | Red flag |
|---|---|---|---|
| 1 | Re-derivable prices | Free public price API, CORS open | Prices only on marketing pages |
| 2 | Failure billing policy | Written refund-on-failure rule | Silence or vague wording |
| 3 | Platform fee | Named, itemized — ideally $0 | "Contact sales" pricing |
| 4 | Supplier depth | Supplier count per model | Single anonymous source |
| 5 | Verified supply | Named merchant + self-check records | Claims without collateral |
| 6 | Quality metrics | Per-lane upstream self-check records | Only aggregate testimonials |
| 7 | Protocol coverage | OpenAI + Anthropic native | One protocol, emulated |
| 8 | Payment + refund | Named channels, refundable balance | Prepaid, non-refundable fine print |
| 9 | Support + incidents | Tickets with SLA + status page | Support = a chat widget |
Everything above is deliberately checkable here: the pricing page mirrors the public API, the catalog shows supplier counts and upstream self-check records per model, the trust page links each promise to its evidence, and /status shows the live platform snapshot. Start with trial credit before your first top-up, then decide with your own workload — the FAQ covers the questions that remain.
The checks take minutes; the decision deserves a loop. Run your real workload on trial credit, read the usage logs against your expectations, and only then move real balance. Nine checks earn you a platform that tells the truth; the loop earns you a platform that fits your traffic. The setup guides get your client connected in the first five minutes of that loop.
Small. Top up the minimum that lets you run your real workload for a day, verify billing behavior and log itemization, then scale. Unconsumed balance is refundable via ticket, but the goal is to spend little while evaluating.
Trial credit lets you call real models before paying anything — enough to validate protocol setup, latency and output quality. Cost behavior at scale is better judged with one small paid top-up.
Supplier depth. Every other failure mode has a workaround except a platform whose whole supply is one lane — when it goes down, your product goes down with it.
In the Terms of Service: failed requests are never billed, pre-deducted quota refunds automatically, and unconsumed balance is refundable via the ticket system (returned by the original payment route, minus payment-channel fees).
Live prices: /en/models/ · billing: /en/pricing · setup: /en/guides/ · FAQ