cocodot
← Back to guides
Local card declined for overseas AI? cocodot: one card + one key
PitfallsUpdated 2026-08

Choosing an LLM API Vendor: a 6-Point Checklist You Can Test Before Paying (2026)

Teams burned by cheap API resellers know the pattern: degraded models, vanishing vendors, billing black boxes. Six verifications you can run in an hour — model authenticity probes, price-equals-charge, no-fee-on-failure — no trust required.

TL;DR: When picking an LLM API vendor, the real enterprise cost isn't the per-token price — it's switching cost and uncertainty: you integrate, then discover degraded models, or the vendor vanishes, or bills don't match list prices. The good news: every one of these risks is testable before you pay. Six checks: ① model authenticity — run an open-source probe (probe.cocodot.co works on any vendor, including us); ② listed price = actual charge — top up the minimum, run 10 calls, reconcile line by line; ③ failures aren't billed — trigger a timeout on purpose; ④ traceable entity — registered company, reachable support; ⑤ sustainable pricing — a price permanently below the provider's official cost is subsidy or substitution, neither durable; ⑥ honest capability boundaries — vendors that document what they DON'T support beat ones claiming everything. Each check takes under an hour. Trust no marketing — including ours.

Why 'cheapest' is usually most expensive

LLM API cost structure is public — upstream prices sit on official sites. A vendor selling below official cost forever can't cover costs; they're either burning subsidy (stops anytime) or substituting the goods (degraded/reverse-engineered/pooled). Integration, migration, and outage costs dwarf the cents saved per million tokens. Buy explainable cheapness, not mysterious cheapness.

The probe: killing the biggest doubt in an hour

Degradation is this industry's trust black hole — you pay for Claude and might get a cheap model in a trench coat. We open-sourced our detector (probe.cocodot.co): six probe families fingerprint the model's actual capabilities, usable against any vendor. Put it in your procurement checklist.

A per-call ledger is table stakes

Finance needs to know where every dollar went. A qualified vendor shows model, token counts, and charge for every call, exportable. A vendor offering only a shrinking balance with no line items can't be reconciled or audited — disqualify immediately.

How we score on our own checklist

Honestly: ① probe is open-source; ② listed prices equal charges, per-call ledger exportable; ③ failed calls never billed; ④ registered overseas company; ⑤ pricing is explainable — no model is priced above its official list, and the discount varies per model (some are clearly cheaper, some match list); check the price page or the models endpoint for current rates rather than assuming one blanket discount; ⑥ our docs list what we don't support: PDF input, no invoices, very high concurrency needs a capacity review upfront, some models carry tiered long-context pricing, and we are a forwarding layer that does not store call content. We learned #6 the hard way — every gap between marketing and implementation eventually becomes a customer incident.

The 6 checks (all pre-payment)

#CheckHowRed flag
1Model authenticityrun the open-source proberefuses testing
2Price = charge10 calls, reconcile eachno per-call ledger
3Fee on failureforce a timeoutfailures billed
4Traceable entitycheck registration, contact supportanonymous payment, TG-only
5Sustainable pricecompare against official costpermanently below cost
6Honest boundariesdoes the doc list unsupported features?claims everything works

FAQ

Isn't direct API always better?

If you can go direct, do. Our value is where you can't: local payment rails, network reachability, many models behind one key. That's checklist #6 — vendors should state their position honestly.

What if this checklist finds a problem with you?

Tell us and we fix it. Price-vs-charge mismatches settle at the lower price. Being auditable is exactly why we open-sourced the probe.

About cocodot

cocodot is a payment and AI access service for developers and cross-border teams in mainland China. It provides US-BIN virtual cards issued by a licensed institution — used to pay for overseas subscriptions and ad accounts — and an OpenAI-compatible AI API gateway for calling Claude, GPT and Gemini from within mainland China. Both share one wallet, funded by Alipay and accounted in USD. Card: $9.9 to open, 3% to load, $1 per active card per month; spending: $0.60 settlement fee on purchases under $20; a corresponding fee applies when the issuer charges one.

Service scope, pricing and limits →