Subscription or pay-per-token — how to find the break-even point

The "subscription or pay-per-token" argument is usually conducted on vibes. In fact it's a single division, and the answer differs per person — it depends on how much you actually work.

The formula

A subscription wins when your monthly token spend would exceed its price. The break-even point:

At $0.30 per million tokens, a $20 subscription breaks even at 66 million tokens a month. A $100 one at 333 million.

Substitute your own subscription price: they change, and a figure from a six-month-old article isn't worth using.

What 66 million tokens looks like

An abstract number, so let's convert it into work.

An ordinary chat conversation runs from a few thousand to a few tens of thousands of tokens. Reaching 66 million by chatting means hundreds of conversations a day, every day. For chat, a subscription almost never pays off.

Agent work is a different story. There, every step ships the whole accumulated conversation, and an hour of dense work easily produces several million tokens. Twenty working days at two or three hours each puts you in the tens of millions.

Hence the practical answer. Chat — almost always tokens. Agent work for many hours daily — run the numbers, a subscription may win. Agent work a few times a week — tokens.

Three factors that break the calculation

Subscription limits. Subscriptions are rarely unlimited: there are windows, quotas, and throttling. Compute against the volume actually permitted, not the headline price — otherwise you're comparing incomparable things. If you hit the cap, your real price per used token is higher than advertised.

Lumpiness. Development comes in bursts: a week of dense work, then two weeks of small fixes. A subscription is paid in full during the quiet weeks; tokens are not. Compare a quarterly average, not your busiest month.

Lock-in to one tool. A subscription usually grants access to one vendor's models. Per-token billing grants access to all of them — and that's not only about price: when one model goes down you switch instead of waiting.

Measuring on your own data

Don't guess — measure. A week on tokens gives you an exact figure, cheaper than any wrong choice.

curl https://api.zukko.pro/v1/usage \
  -H "Authorization: Bearer fetch_your_key"

/v1/usage returns the charge history for the whole account (all keys together): each entry has created_at, model, input_tokens, output_tokens, and cost_cents. The dashboard shows the same spend broken down by model. Multiply the weekly total by four, compare against the subscription price, and the argument is over.

While you're there, look at what share went to simple tasks. Often a noticeable part of the work can move to a model three times cheaper (glm-5.3 at $0.10 against $0.30 per million) — and then the break-even point recedes well past any subscription.

About the price per token

It differs between providers, and that matters more than the subscription question. The official price for Claude Opus 5.5 is $4 per million in and $20 out. At Zukko↗ the same model costs $0.30 for both.

At that gap the subscription break-even moves into volumes almost nobody reaches. Do the arithmetic with your own price, not the official one. One caveat for agent work with a large repeated context: vendors charge about 10% of the input price for prompt-cache reads, while Zukko bills cached tokens at the regular price, so in those scenarios the real difference from official pricing is smaller than a head-to-head comparison suggests.