Cheap LLM API: How to Tell If Cheap Is Actually Cheap

Updated 2026-11-19 · Tapix AI

A cheap LLM API is not the one with the lowest advertised input price. It is the one with the lowest output price on the model you actually route to, after purchase fees, on credit that does not expire, from a provider that does not bill failed requests.

Those four conditions decide the bill. Everything else — banner numbers, "from $0.10/M," marketing comparisons against a model you will never call — is decoration.

This guide gives you a way to check all four in about twenty minutes, using your own usage numbers instead of someone else's benchmark. It is written for developers and small teams who run agents, chat frontends, or batch generation and cannot afford to discover a pricing trap three weeks into production.

Quick answer: a cheap LLM API is one whose effective cost per 1M output tokens on your actual model, including purchase fees and unused-credit risk, is lower than the alternative. Advertised input prices and "starting from" figures are not comparable across providers.

Why the advertised number almost never matches the bill

Three things get lost between the pricing page and the invoice.

The first is which token you are paying for. Every provider prices input and output separately, and output usually costs several times more. A provider can look cheap because it quotes an input rate on the front page.

The second is what the price excludes. A credit purchase fee, a minimum top-up, a monthly platform charge, or a "convenience" surcharge all move real cost without touching the advertised per-token number.

The third is what happens to money you do not spend. Credit that expires is a 100% loss on the unused portion. A discount you must consume within 30 days is a discount only if your usage is predictable, which for most early-stage products it is not.

That is the whole problem. It is not that providers lie about the per-token rate — usually they do not. It is that the per-token rate is one line out of four.

When you compare API providers, compare these line items, in this order:

  1. Output price per 1M tokens, on your real model
  2. Input price per 1M tokens, on your real model
  3. Fees applied at purchase, top-up, or withdrawal
  4. Whether unused credit expires
  5. Whether failed or retried requests are billed

Providers that win on line 1 and lose on lines 3 and 4 are the common case. Providers that win on all five are rare enough that they are worth finding.

The number that actually decides your bill

For most applications built on top of LLM APIs, output tokens are the majority of the spend.

The reason is structural. An agent loop sends a short instruction and receives a long reasoning-and-answer string, then does it again with the answer appended. A chat product sends a growing transcript and receives a reply. A content pipeline sends a brief and receives an article. In each case generation outruns prompting.

The exception is retrieval and long-context work: stuffing a 200-page document into the context window makes input dominate instead. If that is your workload, invert the analysis — but do it consciously, because most teams assume output and are right.

Here is the arithmetic, with placeholder rates. The percentages are the point, not the prices — substitute your own rates from whichever pricing pages you are comparing.

Line itemTokens per monthPlaceholder rateMonthly cost
Input (prompts, context, tool schemas)60,000,000$2.00 / 1M$120
Output (completions, reasoning, tool calls)120,000,000$8.00 / 1M$960
Total180,000,000—$1,080

Output is 89% of that bill from 67% of the tokens. A provider that is 50% cheaper on input and identical on output saves $60 — about 5.5% of the total. A provider that is 20% cheaper on output saves $192 — nearly 18%.

This is why "cheap" is a claim about one number, and it is the number most pricing pages bury.

Comparison trap, same workload: Provider A advertises $1.00 input / $12.00 output. Provider B advertises $3.00 input / $6.00 output. A looks three times cheaper on the headline figure. On this workload A costs $60 + $1,440 = $1,500 and B costs $180 + $720 = $900. The "expensive" one is 40% cheaper. (Placeholder rates, illustrative only.)

Four ways to buy the same model

There are four common routes to a frontier model. They differ less in what the model can do than in where the margin sits.

RouteWhere the price comes fromFees and terms to checkBest for
Model gateway (e.g. Tapix AI)A per-model discount off the vendor's list price; ours is up to 30% below list, shown next to the list price on our pricing pagePrepaid credit, no monthly fee, credit does not expire, $10 minimum top-up, failed requests not billedTeams who want several vendors behind one OpenAI-compatible key and a fixed prepaid budget
Aggregator with purchase fee (e.g. OpenRouter)Inference is billed at the model provider's list price; the platform earns a fee when you buy credit (5.5%, or 8% on its Business plan — check their pricing page for the current figure)Purchase fee at top-up; confirm whether unused credit expiresTeams who want the widest catalog and a well-known brand
Vendor direct (Anthropic, OpenAI, Google, xAI)The vendor's published list pricePostpaid card billing; one account, key and invoice per vendor; enterprise terms, DPAs and support contracts availableAnything that needs a contract, a DPA, procurement paperwork or named support
Consumer-style AI service with API accessVaries by plan and changes oftenWe have could not be verified from public sources these; check the provider's own page before comparingHobby use where the plan bundle matters more than per-token cost

Two honest notes about that table.

First, the gateway row is a discount off list, not a different model. The model names and capabilities are the vendor's; what changes is the invoice. If a provider will not show you the list price beside its own, that is a signal.

Second, direct is not the expensive option for everyone. Direct buys you things a reseller cannot: a signed contract, a data processing agreement, a named support contact, and a vendor relationship that survives an argument. Those have real value and we do not sell them. See when not to switch off direct below.

The three checks

Run these in order. The first one eliminates most of the field.

Check 1: Price the output token, on your model

Open the pricing page. Find the model you actually call in production — not the flagship you demoed with. Read the output price per 1M tokens, not the input price, and not the "starting at" figure.

Then do it again for the second model in your fallback chain, because that is the one that gets called during an incident and it is the one nobody priced.

Fail signal: the page shows one blended number, or only input prices, or prices a model you cannot identify.

Check 2: Add every fee between your card and your tokens

Write down what you pay to put $50 on the account, and what you pay to take it back out. A purchase fee is part of your cost per token — it is just charged once, at the top, instead of per request.

If the fee is 5.5% and you spend the credit down to zero, your effective discount is 5.5% smaller than advertised. If you spend only a third of it before switching providers, the fee lands on the third you used.

Fail signal: fees disclosed only in terms of service, or a "platform fee" that appears at checkout.

Check 3: Ask what happens to credit you do not spend

Two questions. Does unused credit expire? Can you get it back?

An expiring balance converts every unused dollar into a donation. A non-expiring balance is a stored asset — the worst case is that you still have the money when you need it. This single term often matters more than a few percent of per-token discount, because it changes what your prepayment is.

Fail signal: "credits valid for 12 months," silence on the topic, or a policy page that contradicts the pricing page.

A fourth check costs nothing and catches the rest: send a deliberately failing request. A malformed body or an over-length prompt should return an error, not a charge. Providers that meter attempts rather than output punish agent loops, which retry constantly.

How to measure your own number in twenty minutes

  1. Pull last month's usage. Most providers expose per-request token counts. Export a week and multiply by 4.3.
  2. Split input from output. Two columns. If your provider only gives a total, take one representative request and read the usage object directly from the API response.
  3. Apply each candidate's rates to your own split, not to a benchmark workload.
  4. Add fees and subtract the discount you will not consume. If a provider gives 30% off but requires $200/month of usage to keep it, model your real usage, not the requirement.
  5. Compare totals, then compare the failure mode. Two providers within 5% of each other are not a pricing decision — they are a reliability and terms decision.

Do this once and keep the spreadsheet. Rate changes are frequent enough that re-running it quarterly is cheaper than being surprised.

Where "cheap" hides cost

Rolling discounts. A promotion is not a price. If the discount has an end date, model your bill for the month after it ends, because that is the month that hurts.

Expiring credit. Discussed above, and still the most common trap in prepaid products.

Billing attempts. Agent loops retry. If retries are metered, your cost is a function of your error rate as much as your usage.

Silent model substitution. If an endpoint takes a model name and returns output from something cheaper, your unit economics look fine and your product quality quietly degrades. The tell is a pricing page that does not publish the vendor's list price next to its own.

A minimum monthly commitment dressed as a discount. "Spend $500, get 20% off" is not a discount for a team spending $80.

When cheap is the wrong question

Cheap is not the objective. Predictable is.

If you need a DPA, SOC 2, a procurement process, a contractual SLA, or a named escalation contact, buy direct and stop reading pricing pages. A reseller discount does not compensate for a missing contract, and no amount of savings survives a compliance review.

If your monthly spend is large enough that a few percent is real money, negotiate with the vendor or the gateway directly instead of optimising the self-serve rate.

And if your usage is under about $10 a month, the honest answer is that pricing is not your problem yet — pick the provider with the least friction, ship, and revisit when the bill is big enough to argue about.

Where Tapix AI fits

Tapix AI is a model gateway: one OpenAI-compatible endpoint, prepaid credit, per-token billing. Every frontier model is priced up to 30% below the vendor's official list, with the list price shown beside ours so you can check the line instead of trusting the claim. No monthly fee, credit does not expire, $10 minimum, no card at signup, and failed requests are not billed. It is a fit if you want several vendors behind one key and a fixed prepaid budget; it is not a fit if you need a contract, a DPA, or an SLA — we do not offer those.

FAQ

Is a cheap LLM API safe to use in production?

Cheapness is not the risk factor; transparency is. A provider that publishes the vendor list price next to its own, documents its fees, and shows a public status page is checkable. One that quotes a single blended number and buries fees in the terms is not. Run the three checks above before you move production traffic.

Why is output more expensive than input?

Generating tokens is more compute-intensive than reading them, and vendors price accordingly. Because most applications generate more than they consume, output rate dominates the bill. The exception is long-context and retrieval work, where large inputs flip the ratio.

Does a 30% discount mean the model is different?

It should not. A discount taken off the vendor's list price is a pricing decision, not a model change — the model name and capabilities stay the same. The way to verify is to check that the provider publishes the list price next to its own and that the model identifier matches the vendor's.

Is prepaid credit better than a monthly subscription?

For metered usage, usually yes. Prepaid credit does not charge you for capacity you do not use, and non-expiring credit keeps its value if you pause. A subscription only wins if you reliably consume more than the fee is worth in discount.

How do I compare two providers in five minutes?

Take one month of your own usage, split into input and output tokens. Multiply each by each provider's rates. Add purchase fees. Then check whether credit expires and whether retries are billed. Two providers that land within a few percent of each other should be compared on terms, not price.

The short version

Cheap is a property of the invoice, not of the banner. Price the output token on the model you actually call, add the fees between your card and your tokens, and find out what happens to the balance you do not spend. Do that and the "cheap API" question answers itself — usually within one spreadsheet.

If you want to see the per-model comparison for yourself, our pricing page lists every model with the vendor list price beside our price, and the model catalog shows what is currently available. If you would rather compare gateways first, read OpenRouter alternatives compared before you decide.

Output price per 1M tokens: official list vs Tapix AIOutput price per 1M tokens — official list (grey) vs Tapix AI (orange)claude-opus-4-6$25$9.215495 · −57%gpt-6.1-terra$5$1.453417 · −57%claude-fable-5$50$35 · −30%claude-fable-5-1$50$35 · −30%claude-haiku-4-5$5$3.5 · −30%claude-opus-4-7$25$17.5 · −30%claude-opus-4-8$25$17.5 · −30%claude-opus-5$25$17.5 · −30%

Of the 8 models shown, the gap between the vendor's published output price and Tapix AI's price ranges from 30% to 57%, measured on output tokens per 1M. Output is the line that decides most chat and agent bills.

Output price per 1M tokens
ModelOfficial listTapix AIYou save
claude-opus-4-6$25$9.215495−57%
gpt-6.1-terra$5$1.453417−57%
claude-fable-5$50$35−30%
claude-fable-5-1$50$35−30%
claude-haiku-4-5$5$3.5−30%
claude-opus-4-7$25$17.5−30%
claude-opus-4-8$25$17.5−30%
claude-opus-5$25$17.5−30%
See the actual per-model prices. The price page lists the vendor's official price next to ours, row by row, for every model in the live catalog. Open the price list

Related reading

Last updated: 2026-11-21最后更新:2026-11-21