TL;DR (July 2026): A recurring Show HN this year is the "unmetered LLM API" - one flat price (a $6/month example made the rounds), "no token tracking, no limits." It is the mirror image of the credit-metering wave everyone else is riding. These resellers work by buying frontier-model tokens at volume, then reselling flat-rate on the bet that most subscribers under-use, quietly capping the ones who do not through rate limits, concurrency ceilings, model downgrades, and shared-key throttling. Flat pricing is a real convenience for light, hobby, or bursty use where the mental overhead of a meter is not worth it. It is a trap for anything production: "no limits" is never true, "no token tracking" means you cannot see cost or catch abuse, and the reseller can change the terms the moment your usage stops being profitable for them.
The pitch is genuinely appealing after a year of credit cutovers, cap-raising, and token-counting anxiety: pay once, stop thinking about tokens. And for the right user it delivers. But "unmetered" is a business model, not a physics exemption - somebody is still paying per token upstream, which means the flat price only survives if the average subscriber costs the reseller less than they pay. Understanding how that gets enforced tells you exactly when a flat plan will bite.
How a flat-rate reseller actually stays solvent
A reseller selling "unlimited" on top of a metered provider has only a few levers, and they use all of them:
- Gym-membership economics. Most subscribers use a fraction of what they pay for; a minority of heavy users would be unprofitable at the flat price. The whole model is the light majority subsidizing the heavy minority - identical to the wrapper margin, just sold as "unlimited."
- Soft limits dressed as "no limits." Requests per minute, concurrent-request ceilings, daily token soft-caps, and queueing under load. You will not see them until you hit them, usually right when you scale.
- Model routing. "GPT-class" or "Claude-class" can quietly mean a cheaper tier or a quantized open-weight model during peak hours. Flat price, variable model.
- Shared keys and throttling. Many low-cost resellers pool traffic behind shared upstream keys, so one abusive tenant degrades everyone, and the reseller throttles aggressively to protect the pool.
- Terms that move. With no contract-grade SLA, the reseller can add caps, change models, or discontinue the plan the moment your cohort turns unprofitable - which the cheapest ones do routinely.
None of this is necessarily dishonest. It is how any all-you-can-eat offer on a metered input works. The problem is only that the marketing says "no limits" and the reality is "limits you cannot see."
When flat-rate is genuinely the right call
Flat, unmetered access is a good trade in specific cases:
- Hobby and personal projects where predictability beats optimization and $6 is cheaper than the time to think about tokens.
- Bursty, low-average workloads - occasional use that would look expensive on a meter but is trivial in aggregate.
- Non-critical, non-production paths where a rate limit or a model swap is an annoyance, not an incident.
For these, the flat plan removes real friction. This is the same logic behind the free-tier APIs that actually stay free - great for the light case, quietly rate-limited the moment you lean on them.
When "no token tracking" is the actual risk
For anything you run a business on, the missing meter is the buried cost. "No token tracking" sounds like a convenience; it means you have no per-request cost, no per-customer attribution, no abuse signal, and no way to reconcile what you were charged against what you used. If a customer or a runaway agent starts hammering the endpoint, you find out from a rate-limit wall or a discontinued plan, not from a number you control. Production systems need the opposite of unmetered - they need the meter to sit at the gateway where you can set per-key budgets and quotas, precisely so cost and abuse are visible before they become someone else's throttle.
The clean way to think about it: unmetered pricing moves the meter from your side to the reseller's side. That is fine when you do not need to see cost. It is a liability the instant you do - because at that point the only party who can see what is happening is the one whose interest is to cap you.
The honest take
Unmetered flat-rate LLM APIs are a legitimate product for the light, hobby, and bursty cases, and the convenience is real - after a year of credit anxiety, "stop thinking about tokens" is a feature. But treat "no limits, no tracking" as marketing for "limits you cannot see, tracking you do not get." Use them where predictability beats visibility and an occasional throttle is harmless. The moment the workload becomes production - real customers, real revenue, real abuse surface - you want the meter back on your side, because the flat price is only stable until your usage stops being profitable for the reseller, and you will be the last to know.