AI Credits vs Tokens vs Minutes: Unit Economics Guide

In short

Convert each meter to the price of one output you actually produce, then check whether unused units expire. Tokens price text going into and coming out of a language model, with output about five times input on most current models. Speech-to-text bills by the minute of recording, and text-to-speech mostly by the character of script, though some newer speech models bill tokens. Credits are a vendor's own currency that can stand for characters, tokens or seconds of video, so they compare only after that conversion.

On this page4

A 10,000-token request costs 3 cents; a 5-second Gen-4.5 clip 60 cents

Job	Meter	Example rate	Cost of the job
A chat request: 10,000 tokens in, 1,000 out	tokens	GPT-6.1 Sol: $2 per million in, $10 out	$0.03
The same request with the input cached	tokens	cached input $0.10 per million	about $0.011
A 5-minute voiceover, about 4,000 characters	characters	ElevenLabs API: $0.08 per 1,000 (v4), $0.04 (Flash)	$0.32 or $0.16
The same voiceover on an ElevenLabs app plan	credits (1 per character on v2 Multilingual)	Starter: 30,000 credits for $6	4,000 credits, about $0.80 of the plan
Transcribing an hour of audio	minutes	Whisper $0.006 a minute; gpt-transcribe $0.0045; Deepgram Nova-3 $0.0043	$0.36, $0.27 or $0.26
A 5-second video clip	credits (per second)	Runway Gen-4.5: 12 credits a second at $0.01 per API credit	60 credits, $0.60

Visit Leonardo AI

The rates are each vendor's list prices on October 16, 2026; the costs are our arithmetic, and the 4,000-character voiceover assumes about 800 characters a minute, the same figure AI voice API pricing explained uses. The same 4,000 ElevenLabs characters cost $0.32 on the API and about $0.80 of a Starter plan, two and a half times as much.

Output costs five times input; cached input a tenth or less

On most current models output costs about five times input: GPT-6.1 Sol, Claude Sonnet 5.5 and Opus 5.5, and Gemini 3.8 Flash all follow that ratio, while GPT-5.6 Terra charges six times. Caching cuts the input side by 90% or more: OpenAI charges a tenth of the input rate for cached tokens (a twentieth on GPT-6.1 Sol), Anthropic a tenth (a twentieth on Opus 5.5 and Sonnet 5.5), and Gemini 3.8 Flash a tenth. Writing to the cache costs extra, 1.25 times input on OpenAI's GPT-5.6 and later and on Anthropic's 5-minute cache, so caching pays off only when the same prefix is read again.

An ElevenLabs credit is a character; a Runway credit is a twelfth of a Gen-4.5 second

ElevenLabs credits are characters (1 per character on v2 Multilingual, 0.5 to 1 on its Flash and Turbo models), and speech-to-text in its apps costs 330 credits a minute. A Higgsfield credit has no fixed length of video: its rate table charges about 7 credits for a 5-second 720p Kling 3.0 clip and about 60 for a 5-second Seedance 2.0 clip at 1080p (Higgsfield credits explained). A Manus credit also pays for work: Manus's examples run from 200 credits for a 15-minute chart to 900 for an 80-minute web app (Manus credits explained), and Pro starts at $20 for 4,000 a month (Manus pricing). A Lovable credit pays for work rather than output: Lovable's examples put a small Build mode edit at 0.50 to 0.90 credits and a Plan mode message at 1 (Lovable credits explained). Runway's credits are priced per second of video at a model-dependent rate, 12 a second on Gen-4.5 and 5 on Gen-4 Turbo through the API, which is why video plans compare best per kept second.

Runway and ElevenLabs carry some credits over; Deepgram's balance never expires

Plan credits usually reset, but the rules vary even within one vendor. Runway's Standard and Pro credits reset each billing date, its Max plan rolls unused credits over for a month, and credits bought separately never expire. ElevenLabs rolls unused plan credits over for up to two months on paid plans only, with rolled-over credits capped at twice the monthly quota. Deepgram's prepaid pay-as-you-go balance has no expiration.

Official sources

Editorial guidance grounded in official product sources.

Show all 9 sources

Runway documentation

FAQ

Common questions

Is an AI credit the same as an API token?

No. A credit is usually a vendor-specific allowance or wallet used inside a product, while an API token is usually a model-processing unit counted from input, output, cached context, image, audio, or reasoning work. Some products call their internal allowance tokens, so always read the vendor's definition.

Can I compare two tools by the number of credits they include?

Only after you know the burn rate for the exact workflow. Credits can represent different model costs, output durations, resolutions, retry behavior, rollover rules, and feature eligibility. Compare the cost of the same job, not the headline credit count.

Does a paid AI subscription usually include API usage?

Usually no unless the official source says so. App subscriptions commonly cover human access to a product surface, while API usage is billed through a separate developer, project, or organization account. Keep subscription and API budgets separate by default.

What does a minute mean in AI pricing?

It depends on the product. A minute can mean generated video length, audio input length, realtime voice session time, or a feature limit. Check whether the meter is based on input, output, processing time, resolution, audio, or model choice before comparing minute-based claims.

When is seat pricing better than usage-based pricing?

Seat pricing is cleaner when the main need is predictable human access, workspace controls, and admin ownership. Usage-based pricing is better when software, APIs, or variable workloads drive the cost. Many teams need both: seats for people and usage budgets for systems.

What is the fastest way to avoid a bad AI pricing comparison?

Write one sample workflow and map it to each vendor's actual meter: credits, tokens, generated seconds, GPU time, seats, or API usage. Include retries, quality settings, reset rules, top-ups, team pooling, and whether the work happens in the app or through an API.

Next steps

AI Credits vs Tokens vs Minutes: How to Compare AI Tool Pricing Units