LLM API for Captions: What It Costs per 1,000 Videos
An llm api for captions is billed by tokens, and captions are short, so the bill is small: about $0.00012 for a caption written from a short transcript, or $0.12 per 1,000. Below are the exact numbers, how to measure your own, and where the money actually goes.

Key points
- Price is $0.15 per 1M input tokens and $0.50 per 1M output tokens, prepaid, with no subscription and no free tier.
- A caption from a 400-token transcript with a 120-token reply costs $0.00012, so 100,000 captions cost $12.00.
- Input usually dominates the token count, but output costs more per token; trim both.
- The one-time $0.50 trial covers roughly 4,000 captions of that size, and top-ups start at $20.
The short answer
Formula: cost = input tokens × $0.15 / 1,000,000 + output tokens × $0.50 / 1,000,000.
A caption job has two token counts: what you send (instructions plus the text to caption) and what comes back (the caption). Plug both into the formula and you have the price of one item. Multiply by your volume. That is the whole model: no seats, no monthly fee, no minimum usage. You top up credit, you spend it as requests succeed, and failed or refused requests are not charged.
The rest of this page applies the formula to four common caption workloads, then shows how to measure real usage from your own traffic so you can stop estimating.
Four caption workloads, priced exactly
Token counts below are round figures chosen to be realistic, not measurements of any particular tool. Replace them with yours.
| Workload | Input / output tokens | Per item | Per 1,000 | Per 100,000 |
|---|---|---|---|---|
| Short caption from a title and tags | 150 / 40 | $0.0000225 + $0.00002 = $0.0000425 | $0.0425 | $4.25 |
| Caption from a 2-minute transcript | 400 / 120 | $0.00006 + $0.00006 = $0.00012 | $0.12 | $12.00 |
| Description with a 1,500-token style guide | 1,800 / 100 | $0.00027 + $0.00005 = $0.00032 | $0.32 | $32.00 |
| Summary of a 20,000-token transcript | 20,000 / 500 | $0.003 + $0.00025 = $0.00325 | $3.25 | $325.00 |
To check one row by hand: 400 input tokens is 400 / 1,000,000 × $0.15 = $0.00006. 120 output tokens is 120 / 1,000,000 × $0.50 = $0.00006. The sum is $0.00012, and 1,000 of them is $0.12.
Two things stand out. Per-item cost is tiny until the input gets long, and the input side grows fastest when you repeat a big style guide in every request. The third row sends 1,500 tokens of rules to produce a 100-token description, so about 83% of its input tokens are the same text every time. That is where the savings are.
Why output tokens deserve attention
Output costs $0.50 per 1M against $0.15 for input, a ratio of about 3.3 to 1. A caption is mostly output, so it is easy to overspend there by letting the model ramble. Three habits help:
- State a length in the prompt: "one sentence, under 140 characters" works better than hoping.
- Set
max_tokensas a safety cap. It is optional on this API, but a cap of 80 on a caption job stops a runaway reply from costing more than the item is worth. - Ask for the caption only. Without an instruction, models like to add a lead-in sentence and a closing offer to revise. Say "reply with the caption and nothing else".
Cutting a reply from 120 tokens to 60 saves 60 × $0.50 / 1,000,000 = $0.00003 per item. Across 100,000 captions that is $3.00, which is a quarter of the whole $12.00 bill for the second workload.
Measure your own numbers
Every response carries a usage object with prompt and completion token counts, so you do not need to guess. The response headers also include X-Balance-USD, the remaining credit after the request. Log both for a day of real traffic and you have your true average per item.
from openai import OpenAI
import os
client = OpenAI(api_key=os.environ["API_KEY"], base_url="https://api.unfilteredvideo.com/v1")
IN_PRICE, OUT_PRICE = 0.15 / 1_000_000, 0.50 / 1_000_000
def caption(transcript):
raw = client.chat.completions.with_raw_response.create(
model="uncensored",
max_tokens=80,
temperature=0.5,
messages=[
{"role": "system", "content": "Write one caption under 140 characters. Reply with the caption only."},
{"role": "user", "content": transcript},
],
)
resp = raw.parse()
u = resp.usage
cost = u.prompt_tokens * IN_PRICE + u.completion_tokens * OUT_PRICE
print(f"{u.prompt_tokens} in, {u.completion_tokens} out, ${cost:.6f}, balance {raw.headers.get('X-Balance-USD')}")
return resp.choices[0].message.content
Keep a running total in a spreadsheet or a small table, grouped by prompt version. When you change a prompt, the per-item cost changes, and you want to see that before the bill does.
Estimating tokens before you have traffic
Before you have real logs, estimate from text length. For English prose a rough rule is about four characters per token, or three quarters of a word. Other languages and text full of symbols or code use more tokens per word, so treat the rule as a floor. A two-minute talk at roughly 150 spoken words a minute is about 300 words, which lands near 400 tokens. That is where the 400-token row in the table came from.
Then think in months, not requests. Say you publish 30 videos a day and write one caption and one description for each. At the figures above that is $0.00012 + $0.00032 = $0.00044 per video, and 30 × 30 = 900 videos a month costs 900 × $0.00044 = $0.396, so about $0.40 a month. Most teams will find that the caption model is not their cost problem. Retries, oversized prompts and forgotten test loops are, and those show up quickly in the usage log from the section above.
Five ways to lower the bill
Shorten the repeated part
A 1,000-token system prompt sent with 100,000 captions is 100,000 × 1,000 = 100M input tokens, or $15.00. Rewriting it to 300 tokens removes 70M tokens and saves $10.50. Write rules once, in plain words, and cut examples down to the two that matter.
Put several items in one request
If the instructions are long and the items are short, batch them. Twenty separate requests with a 1,000-token prompt and a 100-token item each send 20 × 1,100 = 22,000 input tokens, which is $0.0033. One request with the prompt once and 20 items sends 1,000 + 20 × 100 = 3,000 tokens, which is $0.00045. Ask for a JSON array back (JSON mode is supported) and match results to items by an id field. The trade-off is that one malformed reply costs you the whole group, so keep groups small enough to retry.
Skip the model when you can
If a caption is a template with a title in it, use a template. Send only the items that need judgment: tone, summary, rewording around sensitive material.
Cut input you do not need
Strip timestamps, speaker tags and filler from a transcript before sending it if the caption does not use them. Fewer tokens in is a direct saving.
Do not pay twice for the same job
Store each result keyed by a hash of the input and the prompt version. Re-running a catalogue after a crash should cost nothing for the items already finished.
Trial credit, top-ups and what $20 buys
There is no free tier. New users can claim $0.50 of trial credit with one click on the API key page after signing up with Google or with email and password. It is one per person, valid for 7 days, and needs no card. At the second workload ($0.00012 per item), $0.50 covers $0.50 / $0.00012 = about 4,166 captions. At the 20,000-token summary rate ($0.00325), it covers about 153.
After the trial you top up with any whole amount from $20 to $500, paid with USDT on TRC20 or USDC on Base. At $0.00012 per caption, $20 pays for roughly 166,000 of them. Larger top-ups carry a bonus: +5% from $50, +10% from $100, +15% from $200. A $100 top-up with the 10% bonus gives $110 of credit, which is about 916,000 captions of that size. Credit never expires, so there is no pressure to spend it in a month. The pricing page has the current table.
When this is not the right choice
Be honest about fit. If your captions are plain, safe and high-volume, and you are happy to run a small model on your own hardware, local inference may cost less once you count only electricity. If you need a vendor's official model with a specific brand behind it, this is not that: it is one open-weight model run by an independent provider, reached at a single model ID, uncensored. Check other providers' current pricing before you decide, because it changes.
The case where this API fits is captions and descriptions for material that mainstream APIs sometimes refuse: crime stories, horror, dark humour, mature fiction for adults. The model answers those without content refusals for lawful adult use, and refused requests are not billed anyway. It handles text only, so it will not transcribe audio or render frames. For the larger picture of where it sits in a media stack, see the article on video pipelines, and for request details see the quickstart.
More guides
Questions and answers
How much does a caption cost on this API?
It depends on tokens. A caption written from about 400 input tokens with a 120-token reply costs $0.00012, so 1,000 captions cost $0.12 at $0.15 and $0.50 per 1M tokens.
Is there a free tier for caption jobs?
No. New users can claim a one-time $0.50 trial credit, valid for 7 days with no card. After that, credit comes from prepaid top-ups starting at $20.
Are refused or failed requests charged?
No. Failed or refused requests are not charged, so retries after an error do not add to your bill.
How can I see what each request cost?
Read the usage object in the response for token counts and multiply by the per-token prices. The X-Balance-USD response header shows your remaining credit.
Your key is one form away
Create an account, copy the key, change the base URL. That is the whole setup.