Models

Subscription vs API: how to work out which one is cheaper when the model eats tokens

Every price, credit rule and usage window below was read on the vendor's own page on 19 or 20 November 2026. Three links are referral links: if you use them we may earn a commission or credits, at no extra cost to you. All other links go to the vendor and earn us nothing.

Subscription vs API: how to work out which one is cheaper when the model eats tokens
Subscription vs API: how to work out which one is cheaper when the model eats tokensLalmch / pixabay
By James Calloway8 min read

A subscription is not a cheaper API. It is a prepaid window with an expiry date, and the question worth asking is where your own bill crosses the line where that window starts to pay for itself.

Three vendors now hand out API credit worth exactly the price of the subscription: Claude Max at $100 gives $100 of Claude API credit a month and Ultra at $200 gives $200, and the ChatGPT Pro tiers at $100, $200 and $500 give the same again in API credits. That is parity, not a discount, and parity has a clean break-even point.

The second class sells tokens instead of dollars: Z.ai's GLM Coding Plan starts at $18 a month with a five-hour window and a weekly window that reset independently, and MiniMax M Plan at $22, $55 and $132 meters the same way. There you compare tokens per window, not dollars per month.

Subscription vs API: what each plan really buys

Plan and price What the money buys The limit that bites Who it is for, who it is not
Claude Max and Ultra - $100, $200 a month $100 and $200 of Claude API credit: the API, Managed Agents, the Agent SDK, the playground Credit dies at the cycle end; Claude Code and extra app usage are not covered Traffic past $100 a month. Not a $40 side project
ChatGPT Pro - $100, $200, $500 The same amount again in monthly API credits for your own apps and agents Codex through API keys is not covered; first grant prorated to the cycle OpenAI stacks whose API bill passes $100 a month
Alibaba Cloud Token Plan - Lite $6, Essential $10, Standard $18, Pro $43 Monthly quotas of 20,000, 45,000, 80,000 and 200,000 Credits A quota, not dollars; the international Model Studio edition only Volume on Qwen- and DeepSeek-class models. Pro was $68 until 9 November 2026
Z.ai GLM Coding Plan - from $18 a month Five-hour and weekly credit windows: Lite 2,000/10,000, Pro 12,000/60,000, Max 28,000/140,000 Off-peak is half the credit rate, peak is Mon-Fri 14:00-18:00 Singapore time Coding agents on GLM. At 95-98% cache hit the vendor puts Max at 676-1,463M tokens a week
MiniMax M Plan - Go $22, Explore $55, Build $132 Quota of 1x, 3x, 7.5x against Go, in a five-hour and a weekly window Image and video hit only the weekly window; M Plan key and pay-as-you-go key are not interchangeable Individuals. MiniMax points production at pay-as-you-go
Openference Minion and Boss - $20 and $80 a month $60 and $160 of usage a month at catalog rates: 3x and 2x the fee (our arithmetic) Terms exclude commercial production, background agent loops, batch pipelines and headless daemons Interactive development. Not an unattended fleet

OpenCode Go at $10 a month counts allowances over five-hour, weekly and monthly windows, with per-model monthly limits from $15 to $60 on Go.

Where the plan does not cover your traffic

The exclusions are where a subscription stops being a substitute for an API bill. OpenAI's credits do not pay for Codex through API keys and do not raise Codex limits. Anthropic's cover the Claude API, Managed Agents, the Agent SDK and the playground, not Claude Code. MiniMax splits the two keys, and OpenCode Go refuses any client that does not send a session ID.

The rule that bites hardest is about unattended work. Openference states it in its own terms: subscription allowances are for personal, interactive development, and do not cover commercial software production, background agent loops, batch pipelines or headless server daemons. The same page says it inspects request metadata and rejects non-compliant requests. That is where the human work now sits: a job a person waits for can be priced as a window, a job that runs while nobody is watching has to be priced per token.

Not discounted: what is actually on sale this November

OpenAI and Anthropic published no Black Friday offer. The ChatGPT help-centre index carries no Black Friday article, Claude's help-centre search returns only unrelated hits, and ChatGPT Pro is monthly-only: annual billing and paying for several months in advance are not available.

The one real list-price cut came from Alibaba Cloud, and it is not a Black Friday campaign. An order dated 9 November 2026 cut the international Token Plan Pro from $68 to $43 a month, 37% less, and raised the quota of all four personal tiers. Current Pro subscribers keep the old price for the cycle they are in. A limited new-user price of $3 for Lite, normally $6, and $5 for Essential, normally $10, runs until 23:59:59 on 30 November 2026 (UTC+8) and only for accounts that never bought a Token Plan Personal Edition. GitHub Copilot publishes 30% off Copilot Fusion usage until 30 November 2026, which cuts the model rate and not the subscription, and OpenCode Go halves the rate on Gemini 4 Argon through the same date.

How to work out which one is cheaper

  1. Measure one month of your own list-price bill. That number is your break-even against every credit-for-price plan. A team whose Claude API bill would be $150 a month pays $100 on Max and $50 out of pocket; one whose bill is $40 pays $100 and wastes $60 (our arithmetic).
  2. Compare the same model on the same footing. A credit that funds one vendor's models is not a substitute for a mixed bill. $100 of Claude API credit buys 50 million input tokens of Sonnet 5.5 at $2 per million, or 10 million output tokens at $10 per million (our arithmetic).
  3. Price the cache, not the headline. Sonnet 5.5 lists $2 per million input tokens and $0.10 per million cached reads. Ten million tokens read back from cache cost $1 against $20 uncached (our arithmetic).
  4. Treat a five-hour window as five hours. A window that resets on consumption is not a monthly allowance, and unused quota does not carry over. Continuous work hits that ceiling every week.
  5. Read the terms before you automate the cheap job. MiniMax routes production to a separate pay-as-you-go key, and Openference rejects unattended clients outright. Moving a task to a model is the cheap part; picking a plan that allows it to run unwatched is the decision.
  6. Price the renewal, not the first month. Alibaba's $3 and $5 new-user prices end on 30 November 2026, and MiniMax's 50% launch offer ended on 14 October 2026, so its $22, $55 and $132 are the prices you renew at.

For pay-as-you-go without a dozen vendor contracts, one key for 500-plus models puts 500-plus models behind one key at 0.8 of the official rate on the models it prices, and 1,263 models at list price runs an OpenAI-compatible endpoint at list prices with no deposit fee across 1,263 models.

Frequently asked questions

Is a ChatGPT Pro or Claude Max subscription cheaper than the API? Only above the break-even point, and the break-even point is the price of the plan. Claude Max at $100 gives $100 of credit, so it wins once your own bill passes $100 a month and loses below that. The credit expires at the cycle end, so $60 of traffic under a $100 plan is $40 wasted.

Do API credits roll over into the next month? No. Anthropic states that unused credits expire at the end of each billing cycle with no rollover, OpenAI's programme works the same way, and MiniMax says unused quota does not carry over to the next window or billing period. Only paid-as-you-go balances behave like a bank account.

Can I use a coding plan subscription for an unattended agent? Usually not, and the vendors now say so in writing. Openference's terms exclude background agent loops, batch pipelines and headless daemons, and it rejects requests it judges non-compliant. MiniMax points production workloads at a separate pay-as-you-go key. Read the terms before a cron job depends on a subscription.

What is the cheapest way to run a small agent fleet around the clock? Pay-as-you-go, on one balance, with the models you actually use. A subscription gives you fixed credit that expires and a window that resets; an API bill does not. Holding one key instead of ten is an administrative saving, and for a fleet that never stops it is the only one that survives month end.

Sources

Vendor pages read 19 and 20 November 2026: Claude API credits - Claude API pricing - OpenAI API credits - Alibaba Cloud notice - Z.ai GLM Coding Plan - MiniMax M Plan - OpenCode Go - Openference pricing - Openference terms - CometAPI pricing - NanoGPT pricing

Referral links

Those three links are referral links: if you top up CometAPI, pay for eligible NanoGPT usage or start a Z.ai subscription through them we may earn a commission or credits, at no extra cost to you. Every other link points straight at the vendor and earns us nothing. Prices were correct on 19 and 20 November 2026.

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

791 articles

Related articles

  1. ChatGPT Pro vs Claude Max vs Kimi: which AI subscription is worth it in 2026
  2. OpenAI ships GPT-6.2 Astra and Sol with 20% cheaper API pricing
  3. OpenAI launches $500 Pro 500 tier and cuts limits on $200 Pro plan
  4. OpenAI Reopens $200 Pro Plan, Halves API Credits in Shift to Pay-Per-Use
  5. API aggregators: CometAPI, NanoGPT, OpenRouter and what else, and when one is worth it

« Previous article