Start creating

Leap API

Whisper Large v3 APIOpenAI's model, through Leap

OpenAI's Whisper on fal's fast runtime: the cheapest way to transcribe hours of audio, in 99 languages, and it translates any of them into English. Segment timings, no speaker labels. One Leap key and one balance run it and every other model, priced per call in dollars.

Get an API keyRead the quickstart

First $5 top-up, no card fee.

  • One key for every model
  • Every call quoted first
  • A failed call costs nothing
# Videos and sounds go up in three steps: start an upload of the file's# exact size, PUT the file where the answer says, then complete it.upload() {  UPLOAD=$(jq -n --arg type "$2" --arg name "$1" --argjson bytes "$(wc -c < "$1")" \    '{content_type: $type, bytes: $bytes, filename: $name}' \    | curl -s https://api.tryleap.ai/v1/uploads -H "x-api-key: $LEAP_API_KEY" \      -H "content-type: application/json" -d @-)  echo "$UPLOAD" | jq -r '"url = \"\(.upload_url)\"", "request = \"\(.upload_method)\"",    (.upload_headers | to_entries[] | "header = \"\(.key): \(.value)\"")' \    | curl -sf -K - -T "$1" > /dev/null  curl -s -X POST "https://api.tryleap.ai/v1/uploads/$(echo "$UPLOAD" | jq -r .id)/complete" \    -H "x-api-key: $LEAP_API_KEY" -H "content-type: application/json" -d '{}' | jq -r .id} FILE_ID=$(upload voice.mp3 audio/mpeg) # Then run the model with them.jq -n --arg audio_file "$FILE_ID" '{  "model": "openai/whisper-large-v3",  "input": {    "audio_file": $audio_file  }}' \  | curl https://api.tryleap.ai/v1/generations \    -H "x-api-key: $LEAP_API_KEY" \    -H "content-type: application/json" \    -H "prefer: wait=60" \    -d @- # Video often outlasts the wait and answers "queued" or "running".# Put its "id" in GENERATION_ID and read it again until it is done:# curl https://api.tryleap.ai/v1/generations/$GENERATION_ID -H "x-api-key: $LEAP_API_KEY"

Set LEAP_API_KEY and send it. It uploads your recording first, since inputs take file IDs, not URLs. With prefer: wait=60 the answer holds the result when the call finishes within a minute; a longer one is read again by its ID.

Model ID
openai/whisper-large-v3
Price per call
$0.0027 a minute
This call
It costs $0.0027 a minute.
First top-up
A first top-up of $5, with no fee, pays for your first calls.

The answer

What comes back

A generation. Check status first: succeeded has the file in output, a signed link that works for 24 hours, so download what you keep. failed has error and failure_cause, and costs nothing. usage.cost_usd is what the call was charged.

200 OK, trimmed
{  "id": "gen_...",  "object": "generation",  "model": "openai/whisper-large-v3",  "status": "succeeded",  "output": [    {      "type": "text",      "url": "https://api.tryleap.ai/files/..."    }  ],  "error": null,  "usage": {    "cost_usd": "..."  }}

Inputs

What Whisper Large v3 takes

Every input has a default. The same schema is public, with no key, at api.tryleap.ai/v1/public/models/openai/whisper-large-v3.

Recordingaudio_file
An imageThe sound to transcribe: an uploaded file, by ID. Up to 4 hours.
Videovideo
An imageOr a video whose speech to transcribe: an uploaded file, by ID. Up to 4 hours.
Languagelanguage
Detect, English, Spanish, Portuguese, French, German, Italian, Dutch, Polish, Russian, Ukrainian, Turkish, Arabic, Hebrew, Persian, Hindi, Bengali, Urdu, Tamil, Indonesian, Malay, Filipino, Vietnamese, Thai, Chinese, Japanese, Korean, Swedish, Danish, Norwegian, Finnish, Greek, Czech, Romanian, Hungarian, Swahili, Yoruba. Default: DetectThe language spoken. Detect works for most recordings.
Translate totranslate_to
EnglishSubtitles in another language, timed to the speech. Leave it out to keep the spoken language.

Pricing

Price per call

It costs $0.0027 a minute. A first top-up of $5, with no fee, pays for your first calls. Prices are the provider's list price plus our margin (35% on most models, less on the dearest video), read from the catalog, and a quote gives the exact price of any input for free.

A free quote
# With $FILE_ID from the uploads above.jq -n --arg audio_file "$FILE_ID" '{  "model": "openai/whisper-large-v3",  "input": {    "audio_file": $audio_file  }}' \  | curl https://api.tryleap.ai/v1/quotes \    -H "x-api-key: $LEAP_API_KEY" \    -H "content-type: application/json" \    -d @-

How it works

How the Whisper Large v3 API works

Every model takes the same request, a model ID and its input, and results, prices, failures and limits work the same way for all of them.

  • Wait for it, or hear when it ends

    prefer: wait=60 holds the request until the result is ready, for up to a minute. A longer run is read again by its ID, or a signed webhook tells your server when it ends.

    prefer: wait=60 / "webhook": "https://..."

    In the docs

  • Failed calls cost nothing

    The price is held when a call starts and charged only if it succeeds. A failed or refused call comes back with its failure_cause, and the hold is released.

    "status": "failed" / "failure_cause": "provider_refused"

    In the docs

  • The price before the call

    POST /v1/quotes takes the same body as a generation and returns its exact price, with the input checked, for free. Every result carries what it was charged.

    POST https://api.tryleap.ai/v1/quotes / "cost_usd"

    In the docs

  • Rate limits

    600 requests a minute per API key and 300 per IP address. A 429 carries Retry-After with the seconds to wait.

    RateLimit-Remaining / Retry-After

    In the docs

  • Safe retries

    Send an idempotency key and a retry after a dropped connection returns the same generation instead of starting a second one.

    idempotency-key: <uuid>

    In the docs

  • Batches

    POST /v1/batches starts up to 50 calls in one request, priced and held together, so either all start or none do.

    POST https://api.tryleap.ai/v1/batches

    In the docs

MCP

Or let your agent call it.

Leap's MCP server lets Claude Code, Claude, VS Code and other MCP clients run Whisper Large v3, read its inputs and fetch the result, on your balance. Signing in happens in the browser, so there is no key to paste.

Connect an agentThe agent guide
Add Leap to Claude Code
claude mcp add --transport http leap https://app.tryleap.ai/api/mcp

Claude Code signs you in through the browser the first time it calls Leap.

Teach your coding agent the Leap API
curl -fsSL https://www.tryleap.ai/install.sh | sh

Installs the Leap skill for every coding agent it finds: which endpoint to call, how to read a model's inputs, when to wait and when to poll.

Questions

Whisper Large v3 API questions

How much does the Whisper Large v3 API cost?
It costs $0.0027 a minute. That is the provider's list price plus our margin, the same as in the studio. POST /v1/quotes prices any input for free before you run it.
Can I try the Whisper Large v3 API for free?
No. API keys run on paid credit, so a key's first call needs a top-up, and quotes are free. A first top-up of $5, with no fee, pays for your first calls.
Do I pay for calls that fail?
No. The price is held when a call starts and charged only if it succeeds. A call that fails, or that the model's safety filter refuses, comes back with status "failed" and a failure_cause, and the hold is released.
Is this OpenAI's own API?
No. Leap is a separate service: it sends your call to OpenAI's model and gives you what it makes, so you need no OpenAI account. One Leap key runs Whisper Large v3 and every other model in the catalog, from one balance.
What are the rate limits?
600 requests a minute per API key and 300 a minute per IP address. A batch starts up to 50 calls in one request. A 429 says how long to wait in its Retry-After header.
Can an AI agent call it?
Yes. Leap's MCP server lets Claude Code, Claude, VS Code and other MCP clients find Whisper Large v3, read its inputs, run it and fetch the result on your balance. The client signs you in through the browser, so there is no key to paste.

Build with Whisper Large v3.

Keys run on paid credit, and the first top-up has no fee. The same key runs every model on Leap.

Get an API key