AI companies are competing not just on what each token costs, but on how much useful work a subscription or billing plan buys. A 36Kr report says a comparison of coding subscriptions found that some overseas plans offered substantially more subsidised API-equivalent usage than the domestic plans it examined, challenging the idea that lower sticker prices always mean better value.
Watch Desk analysis
What happened
The report describes a market where companies are using free access, credits, tiered models, API discounts and subscriptions to attract users. It also cites a SemiAnalysis comparison of subscription quotas, calculated against API prices using a particular September usage pattern in which about 96.6% of tokens were cache reads.
Under that calculation, Anthropic’s $200 monthly Max 20x plan equated to about $12,529 of Sonnet 5.5 API usage, or 62.6 times the subscription fee. The report gives lower multiples for the domestic coding plans it compared: 23.4 times for MiniMax Token Ultra, 11.6 times for Zhipu GLM Coding Max and 6.7 times for Kimi Code Ultra. These are estimated API-equivalent values under the chosen usage pattern, not a promise that subscribers can turn their plan into that much cash or get the same value on every task.
Why it matters
A low price per million tokens can be a poor guide to the bill for getting a job done. Subscription quotas, cache pricing, model choice and how efficiently a model completes a task all affect the real cost. The comparison also shows why a single “cheapest model” ranking can mislead: a developer, a casual subscriber and an enterprise buyer are paying for different things.
The report includes anecdotes from developers who say their quotas ran down quickly during coding work. Those experiences illustrate the issue, but do not establish typical usage across customers. The underlying comparison is more useful as a prompt to inspect the assumptions than as a universal league table.
Our read
The useful question is not simply “what does a million tokens cost?” It is “how much of my actual work gets done before I hit the limit?” If you are choosing a plan, compare it on your own workload, including the models available, quota rules and what happens when the allowance runs out. The cheapest headline price may be wearing a very expensive disguise.
What to watch
- Whether providers make quota limits and reset rules easier to compare.
- How the ranking changes for workloads with less cache reading or different models.
- Whether providers publish task-based cost comparisons, rather than relying on token prices alone. The reported comparison is a useful warning against judging AI plans by sticker price alone. We unpack what its figures do, and do not, show, and why the right answer depends on the work you need done. If you were choosing an AI plan, would you trust a standardised price comparison or test it against your own workload first?
Discussion spark: Should AI providers be judged by a standardised cost-per-task test, or is every user’s workload too different for one fair comparison?
Sources and evidence
- Silicon Valley AI Sector Frenzy: How Much Cost-Effectiveness Remains for Players Like DeepSeek? – 36Kr (14 October 2026, 04:02 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.