Is Claude Code cheaper on a subscription or on the API?

It depends entirely on your volume, and you can compute it. The local session logs record token counts regardless of how you are billed, so pricing them at API rates tells you what your actual usage would have cost metered — compare that to your subscription price.

Do subscription plans show a dollar cost?

No. On a Pro or Max plan usage draws against quota, so /cost does not report a dollar figure. There is no per-token charge to report.

Claude Max Plan vs API Billing: Which Is Cheaper?

You already have the data to answer this. Claude Code logs token counts whether you're on a plan or on API billing. Price those tokens at API rates and you have a like-for-like number to compare against your subscription.

npx @tknapp/cli agents --days 30

That total is what last month's usage would have cost on metered API billing. Compare it to what you actually paid.

Reading the result

Equivalent cost well above the plan price → the plan is saving you money. The heavier your usage, the better the plan looks, because a plan is a fixed price for a variable input.

Equivalent cost well below the plan price → you're paying for headroom you don't use. API billing would be cheaper, at the cost of an unbounded downside if usage spikes.

Roughly equal → the plan is still usually the better default, because it caps your worst case. A runaway session on API billing has no natural ceiling; on a plan it exhausts quota and stops. That asymmetry is worth something even at price parity.

Things that make the comparison lie

A month is not representative. Agent usage is spiky — a heavy refactor week is not a normal week. Compare at least two or three months before deciding.

Quota limits are not the same as cost. A plan can be cheaper _and_ still block you at a quota window, which has its own cost in interrupted work. That doesn't appear in a dollar comparison.

Prices move. Anthropic's rates have changed several times, and introductory pricing expires. Any comparison is a snapshot — the numbers TKN uses are dated and source-linked for exactly this reason.

Rate limits differ. Plans and API tiers have different throughput ceilings. If you're throughput-constrained, that may dominate the price question entirely.

The case for API billing

Attribution. Metered usage flows through the billing APIs, so you get per-project cost via organisation reporting. Plan quota does not give you that.
It scales past a plan. No quota window to hit mid-task.
One accounting model if you're already running metered API traffic elsewhere.

The case for a plan

A bounded bill. The single strongest argument. A runaway agent on metered billing can produce a genuinely alarming invoice; on a plan it exhausts quota.
No billing admin. No spend limits to configure, no auto-recharge to remember to turn off.
Cheaper at high usage, usually by a wide margin.

If you're on API billing

Two things to do today, both of which bound your worst case:

Set a hard spend limit — or the Anthropic equivalent — below the number that would ruin your month.
Turn off auto-recharge, so a drained balance actually stops rather than refilling itself.

Then keep an eye on the daily figure. TKN does that across OpenAI, Anthropic, xAI and Gemini, and tokn agents covers the local agent side — different data sources, one place to look.

Sources

Anthropic — Pricing (verified 2026-08-05)
Claude Code — Manage costs effectively