The Real Cost of AI Subscriptions (& Why Pay-as-You-Go Wins)

Subscriptions are a great business model. For the company selling them. They're reliable, predictable revenue that — critically — decouples price from usage. Gyms figured this out in the 1980s. Streaming figured it out in the 2010s. AI companies figured it out last week.

The product you're being sold is access. But the thing you actually consume is compute. And those two things have almost nothing to do with each other.

The gym-membership problem

If you've ever paid for a gym and not gone, you know the shape of this. The gym counts on it. They oversubscribe because they know most people's usage will be a fraction of what they're paying for.

AI subscriptions work the same way. A $20/month ChatGPT Pro subscription gives you a generous allowance — often hundreds of messages a day — because the vendor knows, statistically, you will not come close to using them all.

That's the business. That's the whole trick.

What usage actually looks like

We ran the numbers on Nova's own telemetry (aggregated, anonymized) and the distribution was striking:

• ~60% of users send fewer than 10 messages per day on average.
• ~25% of users send 10–50 messages per day — the steady daily driver cohort.
• ~10% of users send 50–200 messages per day — heavy coders, writers, researchers.
• ~5% of users run agentic or multi-hour flows that push into the thousands of messages per week.

Now translate that to API cost. A short chat message with Claude Sonnet 4.5 costs roughly $0.001–$0.01 depending on context length. A long reasoning session with Opus or o3 can run $0.05–$0.30. Image generations span from pennies (Seedream) to ~$0.15 (premium GPT-5 / FLUX).

If you're in the casual cohort sending 10 messages a day on mid-tier models, your true consumption is maybe $1–$3/month of API value. You're paying $20 for a $3 product because you might want $20 worth someday.

Why subscriptions feel fair even when they aren't

Subscriptions feel fair because they offer predictability. One number, every month, no surprises. There's real psychological value to that — most people will trade measurable waste for the peace of mind of knowing what something costs.

But the moment you look at your actual usage and your actual bill, the illusion breaks. You realize:

1. You paid for the "unlimited" tier but never hit the limits.
2. You have three AI subscriptions and mostly use one.
3. You're paying during months you barely opened the app.
4. When you do hit the limits, they were lower than you thought.

The rational response isn't "use it more to get my money's worth." It's "pay for what I actually use."

The rate-limit trick

Here's the thing the marketing pages don't put in bold type: subscriptions have rate limits, too. They're just disguised.

• ChatGPT Pro throttles Opus-equivalent models to a weekly message budget. Once you blow through it, you're downgraded.
• Claude Pro caps Opus at far fewer messages than Sonnet, and resets on a rolling 5-hour window that's easy to hit mid-session.
• Gemini Advanced has "deep research" quotas, video-generation quotas, image quotas — each with their own ceiling.

So the model isn't actually flat. It's flat within a cap you probably won't see and punishing outside that cap. Pay-as-you-go removes the ceiling: you pay linearly for what you actually consume, no cliffs.

Nova's token model

Nova's answer is a simple credit / token balance that you buy in chunks and spend on any model.

• Every model is priced at its real cost plus a small margin. Cheap models (Seedream, DeepSeek, Haiku-class) are genuinely cheap. Expensive models (Opus, o3, GPT-5) cost what they cost.
• Tokens don't expire on a monthly reset. Buy what you need, use it whenever.
• No auto-billing you forgot about. You buy a balance; when it runs low, you top up.
• One balance covers everything — chat across nine models, image generation across five models, TTS, vision, the works.

The effect: people who chat lightly pay very little (we have active users spending under $3/month). People who go deep on AI pay more but get proportionally more and get access to every frontier model at once. Nobody pays for access they don't use.

When subscriptions still make sense

We're not ideologues. There's a clear case for subscriptions if:

• Your usage is high and steady on one model — Pro often pays for itself at 4+ hours/day.
• You want bundled product features like Sora video, Deep Research, Claude Projects, or Advanced Voice Mode.
• Your employer pays and you'd rather have a fixed-cost expense report line.

For everyone else — which is the bulk of AI users in our data — subscriptions are a structured overpay.

The punchline

For most people, "all the AI you need" costs less than ten dollars a month if you pay by the token. The $20–$60/month subscription stack exists because the default is to sign up for more than you need, and never cancel.

Pay-as-you-go isn't cheaper just because it's cheaper. It's cheaper because it finally lets the price match the work.

Nova starts with 35 free tokens. Buy more when you're ready. Never when you aren't.