TokenCompare

Provider Pitfall Guide

Same models, same workload — real costs can differ several fold between providers. Learn the traps before you pick.

Five traps: where your money quietly leaks

Most of these are invisible on the provider's website — only testing reveals them

01

No caching costs the most

Agent tools like Claude Code / Codex resend the entire conversation each turn. That repeated content is normally billed as cache at ~10% of price. Some providers never enable it — every turn is billed at full price, up to 3x for the same work.

02

Low cache hit rate is even worse

More common: cache is written but never read back next turn — and writing cache itself costs ~1.25x, so you pay on both ends. Independent tests found cases at 7x cost with only 14% actually read back. Hit rates fluctuate and providers don't display them; you just see your balance draining.

03

Model swapping / downgrading

Selling flagship models on the surface while quietly substituting cheaper ones upstream. Cheap providers with murky channel sourcing (subscription slicing, regional discounts, IDE quotas) swap more often. An independent study found 45.83% of relay endpoints failed identity verification — e.g. flagship requests answered by small open-source models, reasoning variants silently downgraded.

04

Billing doesn't match listed prices

One price on the website, another deducted from your balance. Most people only watch the balance drop without auditing what should have been charged. Line-by-line audits have found deductions at 5x the listed price.

05

Exit-scam risk

The owner vanishes with user balances — often within 24 hours, without warning. Nobody can predict who will exit; the only control you have is not parking too much money on one provider. Top up small, more often.

🚩 Red flags

Any single one warrants caution

1Absurdly cheap prices, murky channel sourcing

Prices far below competitors, with the provider unable to explain which upstream channel they use — likely silent model swaps or unstable reverse-engineered routes.

2Weird minimum top-up

A ¥1 minimum means exit-scam costs almost nothing; a ¥100 minimum locks your money into one provider. Both extremes deserve caution.

3No identifiable operator

No company name, registration region, or verifiable contact — if things go wrong, there is nobody to hold accountable.

4Stale for months

Announcements, pricing, and model lists unchanged for half a year — the site is likely abandoned and could go dark anytime.

5No refunds, or harsh refund terms

A provider that pushes all financial risk onto users is a red flag in itself.

✅ Green flags

The more, the more trustworthy

1Transparent channels and pricing

Openly states whether it's official-direct, Kiro, reverse-engineered, etc. — no hiding of cost structure.

2Mid-market pricing

Neither the cheapest on the market (stability sacrificed?) nor noticeably overpriced — mid-tier pricing tends to be honest.

3Welcomes third-party verification

Lets users verify model authenticity with fixed probes or version queries — a sign of confidence.

4Running 6+ months, actively maintained

Survived half a year while keeping pricing and model lists current — a first filter by time itself.

📋 Pre-purchase checklist

Run through these 4 items before you buy

✓Cache verified working, high hit rate
✓Model authenticity check: passed
✓Billing matches listed prices
✓Supports the models you need

🔍 How to test whether a model is real

  1. 1Ask the same math question with a known answer to both the official API and the relay; compare results
  2. 2Ask the model its own version number; vague or evasive answers are a warning sign
  3. 3Compare answer style and length against the official model; noticeable differences warrant suspicion

Used together, these methods rule out most obvious problems but are not 100% conclusive. For an easier route, use our one-click Model Authenticity Check.