Four AI Invoices a Month Taught Me to Audit a Subscription Before Paying for It

The part of AI tooling nobody demos: the invoices

When I moved from writing to video, the hard part wasn't the craft. It was the overhead. One coding agent for drafts and images, a separate voice service for narration, two more platforms for search and clip generation. Four dashboards, four API keys, four invoices — each metered in a different unit, so reconciling them at the end of the month meant converting tokens, characters, seconds and images into the same currency by hand.

I started to suspect my side project had become a bookkeeping exercise.

So when Alibaba Cloud Model Studio upgraded its Token Plan — one subscription, one Credits pool covering text, image, video, speech recognition and realtime voice, plus built-in tools like web search and code interpreter — my first reaction wasn't excitement. It was: does this actually consolidate the mess, or just add a fifth line to it?

Standard and Pro tiers also bundle Harness perks (free monthly quota and discounts on AgentStudio tools), metered separately so they don't draw from Credits. That's the pitch. I have been burned by a plan that looked generous and went half-unused, so I run three audits before subscribing to anything. Here they are, with the actual commands.

Audit 1: the price, and a clock that starts when you do

Official list prices for the personal edition: Lite at CNY 60/month with 2,500 Credits, Standard at CNY 180/month with 10,000, Pro at CNY 600/month with 40,000. Top-up packs are CNY 100 each, up to five held at once. Promo prices show up from time to time; whatever the subscription page charges you is the real number.

The arithmetic on list price works out to roughly CNY 0.024 per Credit on Lite, 0.018 on Standard, 0.015 on Pro — bigger plans are cheaper per unit. That's the part marketing leads with, and it's the part I care about least, because the limit rules decide whether the cheap unit price is real or fictional:

• A fixed 7-day window: the clock starts at your first call of each cycle; hit the cap mid-window and service pauses
• No rollover: unused Credits expire with the window
• Four escape hatches at the cap: buy a top-up pack, upgrade tiers, spend a reset charge, or wait out the seven days
• Concurrency caps: 1–2 on Lite, 3–4 on Standard, 6–8 on Pro

This is a weekly-throughput subscription, not a monthly stockpile. The practical shift in how I read it: I stopped asking "how many Credits do I need per month" and started asking "how many do I burn per week". If your output is steady, the window is invisible. If you publish once a month and spend a whole week's quota in an afternoon, the window will fight you constantly, and no-rollover turns your leftover balance into a loss rather than an asset.

My cadence is two to three videos a week — narration, cover art, the occasional stock clip. Judged against the official Credits deduction table, Standard's weekly allowance covers that with headroom for a sudden topic. That judgment, not the unit price, is what made me consider it.

One scheduling lever worth knowing: a limited-time night window (22:00–08:00) halves Credits consumption on qwen3.8-max and three DeepSeek tiers. Anything that can be deferred effectively costs half, which for batch work is the same thing as doubling your quota. Promo windows change, so check the page rather than trusting my number.

Audit 2: the roster, checked against the live catalog

A plan page listing model names is marketing. A catalog query returning your model ID is evidence. I trust the second one.

The roster spans five families: Qwen (qwen3.8-max, qwen3.8-flash, the 3.7 line, qwen3.6-flash, qwen-image-3.0-pro, and the qwen-audio-3.0 trio for TTS, ASR and realtime), Wan (wan2.7-image and pro), four DeepSeek v4 variants, Zhipu's glm-5.2/5.3, and HappyHorse 1.1 for image-to-video, text-to-video and reference-to-video.

I verified eight of those IDs against the live catalog with the Bailian CLI (bl):

bl model list --model qwen3.8-max --output json

Eight out of eight present, including the two I care about most: qwen-audio-3.0-tts-plus for narration and happyhorse-1.1-t2v for clips. This command needs no authentication, which means you can run the whole audit before you have an account, let alone a subscription.

Then I pulled the plan's TTS voice list — also free information:

bl speech synthesize --config token-plan --model qwen-audio-3.0-tts-plus --list-voices

Two system voices came back, both bilingual Chinese/English. Whether narration is worth subscribing to is a style question, not a spec question, and this is the cheapest way to answer it before paying.

Audit 3: the credential that had quietly expired

Audits one and two passed, so I went to run a real call:

bl text chat --config token-plan --message "Introduce yourself in one sentence"

HTTP 401 InvalidApiKey.

The cause was mundane: the subscription key stored on my machine was from a previous period and had long expired. Nothing warned me. The failure arrived as an auth error, which points at the credential while telling you nothing about which of your stored keys lost the argument.

I also tried to read my Credits consumption directly:

bl usage token-plan

That one needs a console login session, and mine had expired too — so the consumption detail is something I'll read on the subscription page and in the console, not from the CLI on a machine that only holds an API key.

Which means this post contains zero usage impressions. I won't invent them. How natural the TTS actually sounds, whether the image quota survives a real production week — that's the next post, after I refresh the key. Today's post is the audit trail only, and every number in it is either official documentation or something I ran and can show you.

The reusable lesson is smaller than the plan and applies to any subscription: subscription keys and pay-as-you-go keys are different animals with different lifecycles, and expiry sends no notification. Before you pay for a new period, run one cheap call to validate the credential. A 401 means don't pay yet — fix the key, then decide. Keys live on the console API key page, and the region matters: pick China (Beijing).

What the three audits are worth

Running them took maybe twenty minutes and one terminal. What they bought me:

1. A per-Credit number I derived myself instead of one quoted at me, plus the limit rules that decide whether it's meaningful
2. Proof that eight specific model IDs exist in the catalog today, rather than on a launch slide
3. Knowledge that my stored credential was dead — discovered before payment instead of during a production run

The third one is the reason I keep doing this. Every subscription I've regretted failed at the credential or the cadence, never at the feature list.

Who this fits, who should skip it

Subscribe if your output cadence is steady week to week, your toolchain is fragmented across voice/image/search vendors, and you work inside a coding agent — the plan plugs into mainstream coding agents and clients, so it's a key swap rather than a workflow migration, and the learning cost stays near zero.

Skip it if you produce once a month (the weekly window plus no-rollover will waste the balance), if you're optimizing for one model's peak quality (a bundle optimizes for coverage), or if several people would share one seat and your parallel workload exceeds the concurrency cap for that tier. Credits are not the only capacity axis; concurrency is, and it's the one that silently queues your jobs.

If you're on the fence, do what I did: run a shadow month. Take your existing invoices, convert the usage into Credits against the official deduction table, and see which week your real consumption lands in. It costs nothing and it kills the most common mistake — buying a plan that looks generous for the way you work on paper.

Next post: the hands-on bill, in real Credits, once my key is live.