Five Signals Your API Gateway Is About to Fold

"Fine yesterday, today the balance is zero and the group chat is gone." That story repeats every month in the LLM reseller world. But gateways do not vanish overnight: there is a two-week to three-month decline first, and most customers simply never look at the books. All five signals below are measurable, and with a small probe script they become a five-minute weekly check.

Signal one: the balance rules quietly tightened#

This is the earliest and most overlooked sign. The question is not whether there was an announcement, but which direction the terms are moving:

• Balance goes from "never expires" to "valid 90 days"
• Withdrawals close, or a 5%+ fee appears
• Top-up bonuses jump from 3% to 50% or 100%, "limited time"
• Minimum top-up rises from $1 to $20 or $50

The first two are classic liquidity moves. The last two are pulling money in before the exit. A healthy resale margin is 20–40%; a 50% bonus means every sale loses money, and the only rational explanation is that they don't intend to honour it.

Screenshot the terms every time you top up and compare monthly. Terms tighten in one direction only.

Signal two: response quality is drifting#

The most common pre-shutdown move is silent downgrading: your claude-sonnet-5 requests get routed to something cheaper, or shoved into a shared queue. Four measurable indicators:

Fingerprinting sounds exotic but works. Keep three fixed prompts (one coding, one Chinese prose, one math), run at temperature 0 weekly, commit the outputs. When the model behind the endpoint changes, code style and phrasing deform immediately.

Signal three: updates stopped#

How fast a gateway lists new models is its most honest health metric. Doing it requires buying channels, maintaining mapping tables and running regression tests — none of which a team preparing to exit bothers with.

Watch three things:

1. Does the pricing endpoint's syncedAt timestamp still move?
2. How long after a model launch does it appear in the catalogue (normally 3–14 days)?
3. When was the pricing page last updated?

zaimod exposes a public pricing endpoint at https://api.zaimod.com/api/pricing with meta.syncedAt in the payload, which makes a convenient baseline. Any platform with a similar public endpoint can be monitored the same way.

# Check pricing snapshot freshness
curl -s https://api.zaimod.com/api/pricing \
  | node -e "let s='';process.stdin.on('data',d=>s+=d).on('end',()=>{
      const j=JSON.parse(s);
      const days=(Date.now()-new Date(j.meta.syncedAt))/86400000;
      console.log('syncedAt', j.meta.syncedAt, '→', days.toFixed(1), 'days ago');
      console.log('models', j.meta.total, '| priced', j.meta.priced);
    })"

Beyond 14 days stale is worth watching; beyond 30 days, the team has effectively stopped maintaining it.

Signal four: support goes silent#

Support latency is the most sensitive "are humans still here" detector. Check in this order:

• Tickets: does anyone reply within 48 hours?
• Community: when did an admin last speak, and are user questions answered?
• Announcements: how old is the most recent one?
• Status page: is there an independent one, or just pinned chat messages?

The dangerous combination is unread tickets + muted chat + no announcements. Some sites keep an auto-responder alive, but its answer is always "forwarded to engineering" — that counts as a signal too.

Signal five: the payment path changed shape#

Legitimate businesses keep a stable payment path: corporate account, Alipay, WeChat Pay, Stripe. Teams on the way out do three things in the final stretch:

• Rotate receiving accounts frequently (old ones frozen or disputed)
• Accept only USDT or other irreversible crypto
• Use a personal payment QR code whose name doesn't match the site's legal entity

No technical judgement needed — one look at the checkout page. A payee that doesn't match the operating entity is the hardest red flag there is.

A risk scorecard#

Score each signal 0–3 (0 = normal, 3 = already happening), apply weights, read the total:

≤ 2: use normally. 2–5: cap your balance and prepare a backup. > 5: stop topping up, burn the balance, migrate.

The contingency plan: turn a shutdown into a 30-minute migration#

The real risk isn't picking a bad gateway, it's having no exit. Three habits downgrade a shutdown from "outage" to "change one base_url":

1. Externalise base_url — environment variables, never hardcoded. Every mainstream SDK supports it: OPENAI_BASE_URL, base_url=, http_options.
2. Keep balance under one month of usage — top up by usage, never annually, no matter the bonus.
3. Keep a second path — critical flows (payments, support) should never depend on a single provider; use a fallback chain.

Minimal implementation:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["LLM_API_KEY"],
    base_url=os.environ.get("LLM_BASE_URL", "https://api.zaimod.com/v1"),
)

Switching providers becomes a single environment variable, zero code changes. With multiple models it pays off further: one config can point at different vendors, so switching costs almost nothing.

For the seven-point checklist to run at selection time, see how to choose an API gateway; for bill reconciliation, see billing and error codes.

A five-minute weekly probe#

Automate signals two and three; the rest is eyeballing:

#!/usr/bin/env bash
# weekly-check.sh — latency, availability, pricing freshness
set -u
KEY="${ZAIMOD_API_KEY:?missing key}"
URL="https://api.zaimod.com/v1/chat/completions"

for i in 1 2 3; do
  curl -s -o /dev/null -w "run$i ttfb %{time_starttransfer}s total %{time_total}s HTTP %{http_code}\n" \
    "$URL" -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
    -d '{"model":"gemini-3.1-flash-lite-preview","messages":[{"role":"user","content":"ping"}],"max_tokens":8}'
done

Paste three weeks of output into one table. A latency curve trending up and to the right is your signal. Nothing here is platform-specific, so it survives a migration.

Closing#

Shutdowns are visible in advance: terms tightening, quality slipping, updates stalling, people disappearing, payment rails deforming. One signal alone is not fatal; three at once means stop adding money.

Make "can I leave at any time" your first selection criterion — it matters far more than a 20% discount. To check price and latency before committing, pull two models off the pricing page and run the probe above. Data beats promises.

Further reading#

• How to choose an OpenAI API gateway: 2026 checklist
• Gemini API integration in five minutes
• Billing rules and error codes: 429, timeouts, low balance