AI Subscription or API? Choose by How You Actually Use It

Pricing · Published 2026-11-11 · Updated 2026-11-12 · YYSI Editorial · 中文版

The short answer

There is no single "cheaper" — the two products meter different things and buy different things. A consumer subscription (ChatGPT Plus, Claude Pro…) buys interactive features — the chat UI, projects, file uploads, voice — and meters you by messages and time windows with limits the vendors deliberately keep approximate. An API meters tokens per request, is driven by programs rather than a chat window — your own code today, and third-party clients, low-code tools or agent platforms can also call APIs on your behalf where the provider's terms allow — and rewards caching and batching. So the first question is not price at all: how does the work reach the model? If a human clicks through a chat window, a subscription fits. If a program drives the calls — whoever wrote it — the API category fits. Price only decides within the category that fits.

What a subscription actually buys — in the vendors' own words

• Claude (claude.com/pricing, checked 2026-11-11): Free / Pro $20 monthly (or $17 annual) / Max $100 / Ultra $200. Usage limits reset every 5 hours and weekly; the page describes them as multiples (Pro ≥5x Free, Max 5x Pro, Ultra 10x Pro weekly) and says Anthropic "may limit your usage in other ways… at its discretion". Max/Ultra include $100/$200 of monthly API credits; paid plans can also enable overflow usage billed at standard API rates.
• ChatGPT (chatgpt.com/pricing, structure checked 2026-11-11; dollar figures from the YYSI subscriptions hub check of 2026-11-08 because the page renders prices in JavaScript): Free / Go $8 / Plus $20 / Pro from ~$100. The free tier is "unlimited text chats subject to abuse guardrails" and "usage must be reasonable"; Pro tiers come with monthly API credits. No token counts anywhere — by design.
• DeepSeek is the third pattern: the chat app is free and the API is separate pay-as-you-go — there is no consumer subscription to compare against.

What the API actually charges for

API billing is per token, split by direction (input vs output) and by service (fresh vs cached input), with optional discounts for batch processing. Two worked examples verified on official pages on 2026-11-11: DeepSeek deepseek-flash — ¥2 input / ¥8 output per 1M tokens at peak, cache hits ¥0.04, and off-peak (nights, weekends, Chinese holidays) is exactly half; OpenAI gpt-6.1-sol — $2 input / $10 output per 1M, cached input $0.10, Batch API = 50% of standard (but batch results are not real-time, so it only fits workloads that can wait). What the price card does not show: you also pay in engineering time — a client, error handling and retries, rate-limit queues, and someone reading the docs. That cost is real even when the token bill is small.

The arithmetic (reproduce it in our calculator)

Every number below uses the same formula as our AI API cost calculator:

per_request = (Tin x (1-s) x Pin  +  Tin x s x Pcache  +  Tout x Pout) / 1,000,000
per_month   = per_request x calls_per_day x 30

Tin/Tout = input/output tokens per request; s = share of input tokens
served from cache; Pin/Pcache/Pout = per-1M-token prices.

Open the calculator, switch the currency, type the rate row and the workload, and you will get the same figures. The rates below are as-of dated; for the current snapshot always check the live prices-data page or the model-prices JSON rather than this article.

Scenario 1 — light writing (hypothetical)

A solo writer produces 30 short pieces a month, each needing ~2,000 tokens of context in and ~800 tokens out. These volumes are made up; they are typical chat-shaped numbers, not measurements.

Read that honestly: the API bill for a month of light writing is pocket change — around ¥0.31 or $0.36. But this is not an argument that "API is cheaper than the $20 plan". It is an argument that at this volume you never approach what a plan's fair-use limits allow, so the free tier may already cover you, and the paid plan is buying features (projects, uploads, voice, better models), not token volume. If the writing happens in a chat window, the API is not even in the race — you would be paying engineering time to rebuild what the UI already does.

Scenario 2 — batch documents (hypothetical)

A team summarizes 500 documents a month, ~3,000 tokens in and ~500 out each, with a shared 2,000-token instruction prefix that can be cached. A consumer subscription cannot run this at all — there is no programmatic access — so this category is API-only, and the interesting choices are the discounts:

The verdict here is structural, not numeric: at single-digit dollars a month the bill is trivial, and the real decision is which discount fits the workload — Batch only if nobody needs the result in real time; off-peak only for DeepSeek and only for jobs that can run at night; caching only if your requests genuinely share a stable prefix. Note what we did not do: pick a "winner" vendor. The DeepSeek row is in ¥ and the OpenAI row in $; converting between currencies, feature sets and rate limits to declare one universally cheaper is exactly the false precision this guide avoids.

Scenario 3 — app integration (hypothetical)

A product feature calls the model 200 times a day, ~1,500 tokens in and ~600 out each, with a 1,024-token shared system prompt that can be cached — 1,024 being the smallest prefix OpenAI documents as cacheable for its GPT-5.6+ models (68% of the input here). Again API-only — a consumer chat plan cannot serve other people's requests:

Volume dominates everything: doubling traffic doubles the bill, and no subscription is involved at any point. The big lever is not the vendor — it is the model tier: the same workload on a cheap-tier model costs about 5% of the flagship-tier bill. Caching trims another ~20–26% (the exact figure depends on the vendor's read rate). Rate limits, uptime and error handling now matter more than unit price, because an outage is visible to your users, not just to your wallet. One honesty note: the first version of this scenario cached 60% of the input (900 tokens), which is below OpenAI's documented 1,024-token minimum — the earlier "$43.74 with cache" figure was arithmetic on a prefix OpenAI's cache would not have served, and was corrected on 2026-11-12.

Why there is no breakeven point in this guide

• Consumer limits are officially non-token: ratios (5x Pro), time windows (5 hours/weekly), and discretionary caps. A breakeven against "$20" would require inventing the token count.
• The feature sets barely overlap: projects, file analysis, voice and mobile apps on one side; automation, batch discounts and any programming language on the other. Whichever you pick, you are not buying the other side's benefits at any price.
• The boundary is deliberately blurry and moving: Claude Max/Ultra include monthly API credits, Claude paid plans can overflow at API rates, ChatGPT Pro includes monthly API credits. A precise crossover computed today misleads the moment a vendor adjusts a cap.
• Vendors can and do adjust limits at their discretion — their own pages say so.

A checklist instead of a verdict

1. Does a human drive every request? → subscription category; compare feature sets, not tokens. If free-tier limits already cover you, the paid decision is purely about features.
2. Does a script or pipeline drive requests? → API category; compute with the calculator, then optimize: caching for shared prefixes, Batch for deferable work, off-peak for DeepSeek night jobs, cheaper tiers for easy tasks.
3. Does an app serve other people? → API only; budget by volume, add rate-limit handling, and re-check the bill monthly at first.
4. In between (a power user hitting plan caps)? → look at the official bridges: overflow usage billed at API rates (Claude paid plans) or plan-included API credits — priced against your actual overflow, not against a fantasy token quota.

Rates used, and where to check current ones

This article intentionally contains no copy of the full price table — prices move. The rows used above, each with its check date: DeepSeek deepseek-flash ¥2/¥8 peak, cache ¥0.04, off-peak half (official page checked 2026-11-11; hub row 2026-11-08); OpenAI gpt-6.1-sol $2/$10, cached input $0.10, Batch 50% (official page checked 2026-11-11); OpenAI gpt-6-luna $0.10/$0.50 (checked 2026-11-11); Claude Sonnet 5.5 $2/$10 (hub row checked 2026-11-09); consumer plans as quoted in the vendor section (Claude page checked 2026-11-11; ChatGPT structure checked 2026-11-11, amounts per hub check 2026-11-08). For the current full picture use the subscriptions & price comparison hub, the prices-data page with its snapshot history, or the model-prices JSON API — and re-run the calculator with fresh numbers when you decide.