The $300 that never deducts: why you can't tell what your AI actually costs
A developer posted a small, honest question on r/googlecloud. They had a $300 free trial and a separate $1,000 GenAI credit, and they wanted to point them at the Claude model. The credits, they said, “never deduct for anything.” A second person chimed in: “I also got it, but couldn’t spend it on any of my use cases.”
The answers were blunt. You can’t use the free-trial credits toward the Claude model. As one commenter explained it, Claude on Google Cloud runs through a separate click-through agreement between you and Anthropic, transacted via the marketplace — and free-trial credits generally only apply to Google’s own first-party SKUs.
That confusion is the story. Not “how do I redeem a coupon,” but: a working engineer could not tell what their AI usage costs, or whether it was costing anything at all. That’s the same meter you can’t see this whole site is about — except in AI it comes pre-fogged, and the fog looks like a gift.
The short version: three meters, and the free credit only touches one
On any hyperscaler, an “AI” charge can land in one of three billing buckets, and they do not share a wallet:
- First-party SKUs — the cloud’s own services and models (Compute, Storage, Vertex’s own Gemini models). This is where free-trial credits, promotional credits, and committed-use discounts apply.
- Marketplace / third-party models — Claude via Vertex/Model Garden, and other partner models. These are billed under a separate agreement with the model vendor, merely transacted through the cloud’s marketplace. They are typically excluded from free-trial credits and often from your committed-use discounts too. This is the exact wall the Reddit poster hit.
- Direct vendor billing — calling Anthropic or OpenAI directly, outside the cloud entirely. A fourth wallet, with its own separate trial credits and its own console.
So the “$1,000 that never deducts” usually isn’t a bug — it’s a credit fenced to SKUs you’re not actually using. It sits in the console looking like money you have, while your real usage is billing against a bucket the credit was never allowed to touch. (Terms differ per provider and change often — the durable move is to open Billing → cost breakdown by SKU/service and read which line your model calls actually land on, rather than trusting the big “credits remaining” number.)
That’s the helpful, boring answer. Here’s the part that matters more.
A credit that never deducts is a taped-over fuel gauge
We treat free credits as a discount. For understanding your cost, they’re worse than paying full price.
When the meter reads $0 no matter what you do, you get no signal. You can’t tell a cheap prompt from an expensive one. You can’t tell that your retry logic doubled your token spend, or that switching to a bigger model 5בd your per-call cost, because every call still reads zero. You’re driving with the fuel gauge taped over, learning your consumption habits on someone else’s fuel — right up until the tank is yours.
This is the $0-per-call trap from Day 1 with a new coat of paint. Day 1’s version was “the call feels free because it’s a fraction of a cent.” This version is “the call reads free because a credit (that may not even apply) is masking it.” Same outcome: no habit of watching the number, formed at exactly the moment you should be forming it.
Don't confuse a subsidized meter with a cheap one.
”Never deducts” today is “deducts a lot” later
Now zoom out to why the credits exist at all. Every AI vendor and every cloud is currently racing to put trial credits in developers’ hands — $300 here, $1,000 there, generous free tiers, loss-leader token rates. It’s worth asking whether that flood is a bubble about to pop.
I’d frame it more usefully than a prediction: it’s a subsidy, not a discount. This is textbook platform land-grab — the same play the clouds ran for a decade handing out compute credits to lock in switching costs (your data, your integrations, your team’s muscle memory). The AI version locks in your prompts, your eval suites, your fine-tunes, your gateway wiring. Whether or not anything “bursts,” the economics are the same: today’s price is venture-subsidized and below what it costs to serve you.
Which means the dangerous move is to internalize the subsidized number as the number. “Never deducts” is not a price — it’s a promotional window. When the credit expires, the free tier tightens, or list prices normalize to cover the GPUs, the line that read $0 starts reading a real figure. If you built your architecture, your margins, and your habits on the fog, that day is a nasty one. Don’t confuse a subsidized meter with a cheap one.
What to actually do
The fix is the same discipline as the rest of this series — just applied before the money is real:
- Instrument your true cost now, at list price. Route calls through a gateway that logs tokens-in, tokens-out, and the list-price dollar cost of every call — regardless of who’s currently footing it. Your dashboard should show what this would cost with zero credits. That number is the truth; the “credits remaining” number is theater.
- Cap as if the credits weren’t there. Set your hard spend cap and per-user quotas against real projected spend, not the padded free-trial balance. A cap calibrated to “we have $1,300 of credit” evaporates the day the credit does.
- Know which wallet each call hits. Before you commit to Claude-on-Vertex vs. Gemini-first-party vs. Anthropic-direct, check which of the three buckets it bills to and whether your credits and discounts actually apply. Sometimes the “free” path is the most expensive one once the trial ends.
Do that, and the day the subsidy ends is a non-event: the number on your dashboard was always the real one, and your caps were always set to it. The fog was never hiding a discount. It was hiding your uncapped meter — and the tape was going to come off eventually.
Building on a stack of trial credits and not sure what it costs at list price — or which of the three wallets each call is actually hitting? That’s exactly the fog I clear. Every message comes straight to me — I read and reply to each one myself, usually within a day, and what readers send shapes what I build next. It’s just me for now, so that’s genuinely true; it won’t be forever. Send me your AI setup and I’ll tell you what it really costs with the credits stripped out — free, within a business day.