DeepFrugal Early Preview Find the cheapest way to run any model

How token credits work in AI plans

A token-credit AI plan counts usage in the plan's own credits, not dollars. What a credit buys, how multipliers work, and why the value can be a range.

17 min read DeepFrugal
How token credits work in AI plans

A token-credit plan does not count its monthly wallet in dollars. It counts it in the plan's own credits, and each model spends a different number per token. This guide explains how those plans work β€” the wallet, the multipliers, and how far a small fee really goes.

One of three kinds of AI credits

  • Dollar credit β€” the wallet is a budget of dollars, spent at the models' listed rates. How AI credits work covers that case.
  • Token credit β€” the wallet is the plan's own credits, and each model spends a different number per token. This is the case below.
  • Usage limited β€” messages, tasks or a shared pool, with no token or credit amount published at all.

Everything below applies to token-credit plans.

What a token credit is

A plan grants one wallet of credits for the month. Each model has a multiplier: the credits one million of its tokens costs. The wallet divided by the multiplier gives tokens; the model's listed price turns those tokens into dollars. A high multiplier gives few tokens, a low one gives many.

Plan A Β· one wallet of credits12,000 credits/month
Model 1 200 cr/1M 60M tokens
Model 2 400 cr/1M 30M tokens
Model 3 800 cr/1M 15M tokens

Plan A costs $10/month and grants one wallet of 12,000 credits. Each model has a multiplier: the credits one million of its tokens costs. The one wallet buys 60M tokens on Model 1, 30M tokens on Model 2 and 15M tokens on Model 3 β€” the same credits, restated at each model's rate.

Sample plans and models β€” figures are illustrative. Plan A is a sample: a $10/month token-credit plan.

How the multiplier works

The multiplier is not one number. It is set per token type β€” input, cached and output β€” so the same wallet buys a different amount of each. A higher multiplier spends more credits for the same tokens, so it draws the wallet down faster.

Model 1 Β· multiplier by token typecredits per 1M tokens
Input 200 cr/1M 60M tokens
Cached read 300 cr/1M 40M tokens
Output 400 cr/1M 30M tokens

The multiplier is not one number. Model 1 charges a different rate for each token type, so the one wallet buys 60M tokens of input, 40M tokens of cached and 30M tokens of output. A higher multiplier spends more credits for the same tokens, so the wallet buys fewer of them.

Sample plans and models β€” figures are illustrative. The multiplier is the plan’s own rate; it need not track the listed price.

Do the model quotas add up?

No. The plan grants one wallet, not one per model. Each model's token quota is the same wallet restated at that model's multiplier, so adding the quotas together counts the plan several times. What matters is how much of the one wallet your mix spends.

Token quotas added βœ—
Model 1 Β· 60M tokens
Model 2 Β· 30M tokens
Model 3 Β· 15M tokens

60M tokens + 30M tokens + 15M tokens = 105M tokens. Adding the quotas counts the same wallet three times.

One shared wallet βœ“
50% 30% 20%
Model 1 Model 2 Model 3

The same usage as a share of the one wallet: 50% + 30% + 20% = 100%. The whole $10 fee buys this mix.

Sample plans and models β€” figures are illustrative. The per-model token quotas are three views of one 12,000-credit wallet.

How much can the plan really give you?

The most a plan can cover is its biggest single-model quota β€” the listed usage you get if you spend the whole wallet on one model. Split the wallet across models with smaller quotas and you cover less, even when the wallet is fully used. The mix decides.

When a plan's multipliers are not proportional to its listed prices, the same fee buys a different amount of listed usage per token type, so value per dollar is a range, not a point. When the multipliers do track the prices, the value is close to a single number.

Plan A Β· value per dollar$10/month
All on Model 1 $60 6.0Γ— the whole wallet on the biggest token quota β€” the best case
Sample mix $42 4.2Γ— three models share the wallet β€” the usual case
Model 1, by token type $20–$60 2.0×–6.0Γ— the multipliers do not track the listed prices β€” a range, not a point

Value per dollar is the listed usage the wallet covers, divided by the fee. Spending everything on the biggest token quota reaches the best case: $60 Γ· $10 = 6.0Γ—. The sample mix covers $42, so 4.2Γ—. When the multipliers are not proportional to the listed prices, the same fee buys a different amount per token type, and the value is a range.

Sample plans and models β€” figures are illustrative. A token-credit plan can give less than its best case; the mix and the token type decide.

How the wallet is spent

The plan spends the one wallet on the models in the order listed. Each model takes its share, and never more than what is left, so the running total stops at the full wallet. The first example fits inside it; the second runs past it, and the extra usage is paid outside the plan at the listed rate.

Plan A spends the one credit wallet on the models in the order listed, until it reaches 12,000 credits. Each model's credits are usage Γ— multiplier. Running is the Share column added up to that row, capped at 100%.

Example 1 β€” usage within the walletno pay as you go
Model Usage Credits Share = credits Γ· wallet Running Pay as you go
Model 1 30M tokensΓ— 200 cr/1M 6,000 cr 50% 50% β€”
Model 2 9M tokensΓ— 400 cr/1M 3,600 cr 30% 80% β€”
Model 3 3M tokensΓ— 800 cr/1M 2,400 cr 20% 100% β€”
Total β€” 12,000 cr 100% 100% $0

The wallet covers $42 of the usage β†’ 4.2Γ— the $10 fee.

50% + 30% + 20% = 100%. Every dollar of usage is covered by the plan.

Example 2 β€” usage beyond the wallet$9 paid outside the plan
Model Usage Credits Share = credits Γ· wallet Running Pay as you go
Model 1 30M tokensΓ— 200 cr/1M 6,000 cr 50% 50% β€”
Model 2 9M tokensΓ— 400 cr/1M 3,600 cr 30% 80% β€”
Model 3 12M tokensΓ— 800 cr/1M 9,600 cr 80% 100% $9
Total β€” 19,200 cr 160% 100% $9

The wallet covers $42 of the usage β†’ 4.2Γ— the $10 fee, plus $9 paid as you go.

Model 3 asks for 80% of the wallet, but only 20% is left. It covers $3 and the other $9 is paid outside the plan. The running total still stops at 100%.

Sample plans and models β€” figures are illustrative.

Why the cost per token falls, then rises

Read the curve by its average, not by the next token. Inside the wallet the fixed fee spreads over more and more tokens, so the average falls. It reaches its lowest point at the limit of the wallet. Above the limit the extra tokens are paid outside the plan at the listed rate, so the average climbs back towards it. On a token-credit plan the limit is the tokens the wallet buys at the model's multiplier.

1 quota limit listed rate effective rate β€” the lowest average monthly tokens β†’ avg $/1M
  • 1 Quota usage break-even β€” where the average falls to the plan's listed rate. To its left you overpay: the commitment is not recovered.
  • average cost per token listed rate quota limit

Sample plans and models β€” figures are illustrative.

Token credits vs pay-as-you-go

Compare the curve with a flat pay-as-you-go rate β€” the price you pay when you buy tokens one by one. It crosses twice. The first crossing is where the subscription starts to beat the cheapest pay-as-you-go route. The second is where it loses again, because the extra usage above the wallet costs more than paying per token. In between is the sweet spot: the plan covers everything and the average is still falling. That is where the fee gives the most value.

123 quota limit listed rate cheapest pay-as-you-go sweet spot plan + listed still wins monthly tokens β†’ avg $/1M
  • 1 Quota usage break-even β€” where the average meets the plan's listed rate.
  • 2 Plan break-even β€” from here the subscription beats the cheapest pay-as-you-go rate.
  • 3 Loses to the cheapest β€” beyond it the pay-as-you-go route is cheaper again.
  • Sweet spot β€” from the plan break-even to the quota limit: the plan covers everything and the average is still falling.
  • Plan + listed still wins β€” above the quota the plan still wins, but only while the extra usage is paid at the listed rate. The third marker ends that.
  • average cost per token cheapest pay-as-you-go sweet spot quota limit

Sample plans and models β€” figures are illustrative.

Peak and off-peak multipliers

Some plans charge fewer credits at quiet hours. A window like that does not change the listed price; it changes how fast the wallet is spent, so the same fee stretches over more usage. It matters most for usage that comes in short, heavy bursts.

Plan A Β· peak vs off-peaksame wallet, different multiplier
Peak 200 cr/1M 60M tokens
Off-peak 100 cr/1M 120M tokens

Some plans charge fewer credits at quiet hours. At half the multiplier the same wallet buys twice the usage: 120M tokens instead of 60M tokens. The window changes how fast the wallet is spent, not the listed price.

Sample plans and models β€” figures are illustrative. Sample window; the real one is published per plan.

See it on a real plan

The figures above use sample numbers. Here is the same rule on a live plan. The table below lists each model's multiplier β€” the credits one million tokens costs, per token type β€” and the wallet the plan grants. The rates table then shows what those credits buy at the listed prices.

Credit rate per 1M tokens β€” GLM Coding Lite
ModelInput /1MCached read /1MOutput /1M
GLM 5.36901702,400
GLM 5.3 Flash23056800
Credits one million tokens costs, by token type. The plan grants 43,480 credits a month for $18.00; divide the wallet by a multiplier to get the tokens it buys.
Listed vs effective rates β€” GLM Coding Lite
Subscription $18.00/month Β· effective rates are published per token type (quota: 43,480 credits) Β· GLM 5.3 Flash: 230 input Β· 56 cached Β· 800 output credits per 1M tokens
Input /1MOutput /1MCached read /1M
$0.095$0.15$0.331$0.5$0.023$0.03

Browse every AI subscription plan in the live table, or learn how to find the best plan for your usage.

Run your numbers

Put your own mix against a plan. Pick a plan, add models and set the monthly usage. The calculator spends the one wallet in the order shown, marks the break-evens and states the cheaper side.

Plan break-even β€” GLM Coding Lite Β· Experimental
Cheaper
Subscriptionsave $4.00/month
You pay $18.00/month instead of $22.00 of usage.
Usage: 89M tokens Β· $43.40 at list rates Β· $22.00 cheapest pay-as-you-go.
Subscriptioncheaper$18.00/month Β· 43,480 credits
Quota used73%
Credits consumed31,730 cr
Covered by the quota$43.40 Β· 31,730 cr
Above the quota$0.00 Β· 0 cr
Monthly cost$18.00
Cheapest pay-as-you-go route

Real rates include the service fee; sales tax is not included.

GLM 5.3 Flash via LLM Gateway API

GLM 5.3 via OpenRouter API

Monthly cost$22.00
Break-even: $18.00 of usage (β‰ˆ 36.9M tokens at this mix)
At your mix: subscription $18.00/mo Β· cheapest metered $22.00/mo Β· at list rates $43.40/mo β€” subscription is cheapest by $4.00/mo.
Plan break-even details
$0.148$0.202$0.247$0.48850M100M200M89M36.9M72.8M151M122MTotal tokens per month (M tokens/mo, log scale)Average cost per 1M tokens usage above the plan paid at the listed rate your mix
  • Average cost per token
  • Listed rate
  • Cheapest pay-as-you-go route
  • Quota limit
  • Cheaper than the cheapest
  • Quota usage break-even
  • Plan break-even / loses
  • Average at your mix
Hover the curve to read the values.
β–Ύ How the graph is calculated

Inside the quota the average is the monthly commitment divided by the tokens used, so it falls as the mix scales up and reaches the mix's effective rate at the quota limit. Above the limit the pool is spent and the extra tokens are paid outside the plan at the overflow base chosen in Options β€” the model's listed rate (default), the cheapest pay-as-you-go rate found for it, or the cheaper of the two β€” so the average climbs towards that blended rate.

The pool is one shared budget: each model's allowance is that same pool restated at its rates β€” dollars for a dollar-credit plan, credits for a token-credit plan β€” and it is allocated in the model order shown, so a different order or blend moves the curve. A token type a model does not price still counts in the volume but adds no cost. The token axis is logarithmic, so a wide range of usage fits on one chart.

The pay-as-you-go line is the cheapest real rate per model (fee included; sales tax excluded) β€” an exact variant match when a pay-as-you-go gateway publishes that variant, otherwise the model's nearest published row β€” so it can combine more than one gateway. When that rate equals the plan's own listed rate the chart shows the listed line only.

The markers: the quota usage break-even is where the average meets the plan's list rate, so below it you pay more per token than list; the plan break-even is where the average starts to beat the cheaper of the two references; loses to the cheapest is where it stops doing so (it exists only when the overflow base is dearer than that reference); the your mix point is your stated usage, labelled with its monthly token volume, where the horizontal guide marks the mix's average cost per 1M, and the shaded band is where the subscription wins.

β–Ύ How the plan quota is consumed

A subscription grants one monthly pool. For this plan the pool is 43,480 credits; each model spends credits at its own rate (credits per 1M tokens). The pool is allocated to the models in the order listed here: each model spends its usage divided by its quota, capped by what is left, so the total never exceeds 100%. Reorder the models to approximate your own pattern. Usage above a model's allowance, or beyond the pool, is paid outside the plan at the listed rate.

ModelUsageMonthly pool% of QuotaPay as you go
GLM 5.3 Flash12,680 cr$7.90 listed43,480 credits29.2%β€”
GLM 5.319,050 cr$35.50 listed43,480 credits43.8%β€”
Total31,730 cr$43.40 listedβ€”73%$0.00

Pool used = 31,730 cr of 43,480 credits (73%) Β· Consumed = 31,730 cr Β· Covered = 31,730 cr ($43.40) Β· Paid outside = 0 cr ($0.00)

β–Ύ Effective rates (listed vs effective)
ModelInput /1MOutput /1MCached read /1M
GLM 5.3 Flash$0.062$0.15$0.207$0.5$0.012$0.03
GLM 5.3$0.581$1.4$1.825$4.4$0.108$0.26

Effective = listed Γ— 0.415 at this usage Β· 73% of the plan budget used

Experimental. Confirm every figure against the provider's own pricing before you rely on it. DeepFrugal is not responsible for calculation errors.

Not sure how to read it? How the break-even calculator works.

Plans, promos and prices change over time. Check the live table for current values.

Compare live AI subscription prices

Every widget links back to the live comparison table, where you can filter by plan, price and privacy, and sort the results. Four tools cover the decision from different angles:

Related reading: how AI credits work covers the concept and the dollar case, and subscription vs pay-as-you-go walks through the calculator.

Summary

A token-credit plan is one wallet of credits, spent at each model's multiplier. Its best case is the biggest single-model quota; a mixed workload usually covers less. When the multipliers do not track the listed prices, value per dollar is a range, not a guarantee. Estimate your monthly tokens, then check the live table or run the break-even calculator.

Frequently asked questions

What is a token-credit AI plan?

A subscription whose monthly wallet is counted in the plan's own credits instead of dollars. You pay one fee and spend the credits during the month.

What is a token credit?

The plan's own unit of account. The plan grants one wallet of credits, and each model spends a number of credits per million tokens at its own rate.

Is a credit the same as a token?

No. A credit is the plan's accounting unit; a token is text the model reads or writes. The multiplier is the bridge: it says how many credits one million of a model's tokens costs.

What is a multiplier?

The plan's rate for a model: credits per million tokens, set per token type (input, cached and output). A higher multiplier spends the wallet faster.

How many tokens does a credit buy?

It depends on the model and the token type. Divide the wallet by the model's multiplier to get tokens; the listed price turns those tokens into dollars.

Do the per-model quotas add up?

No. The plan grants one wallet, not one per model. Each model's token quota is that same wallet restated at the model's multiplier, so adding them counts the plan several times.

Why can the value per dollar be a range?

When a plan's multipliers do not track its listed prices, the same fee buys a different amount of listed usage per token type, so value per dollar is a range. When the multipliers do track the prices, the value is close to a single number.

How is the wallet spent?

On the models in the order listed, each taking its share, and never more than what is left. Usage above the wallet is paid outside the plan at the listed rate.

Why does the cost per token fall, then rise?

The fee is fixed and spreads over more tokens as you use more, so the average falls to its lowest point at the limit of the wallet. Above it extra tokens are paid outside the plan, so the average climbs again.

When does a token-credit plan beat paying per token?

In the middle of the curve. Use too little and the fee is not recovered; use too much and the extra usage costs more than buying tokens directly. The sweet spot is between those two points and the wallet limit.

Do off-peak hours change the price?

No. A quiet-hours window charges fewer credits, so the wallet stretches over more usage. It changes how fast the wallet is spent, not the listed price.

How do I check a real plan?

Estimate your monthly tokens, then open the break-even calculator and read where you land against the break-evens. It values a plan against the cheapest pay-as-you-go route. The step-by-step tutorial shows how to read it.