Whisper API pricing in 2026: cheaper than OpenAI's official one

May 5, 2026 · 5 min read

If you've looked at OpenAI's transcription pricing recently, you saw $0.006/min for Whisper. That's almost 10× cheaper than LessRec's $0.05/min. So why does LessRec exist?

Because $0.006/min is the price of the model API call. It's not the price of getting a usable transcript. Once you account for what's actually included, the math flips for most use cases.

The all-in cost comparison

Imagine you're transcribing 1 hour of podcast audio (~60 minutes, ~50 MB MP3).

OpenAI is unambiguously cheapest per minute of API time. LessRec is 8× more. But "API time" is the cheap part of running a transcription service. Here's everything OpenAI doesn't include:

What OpenAI's $0.006/min skips

1. The 25 MB upload cap

OpenAI's Whisper endpoint caps each request at 25 MB. A 1-hour podcast at standard MP3 bitrate is ~50 MB — over the limit. Workarounds: compress to 64 kbps, or split the file. Both are doable but they're your code to write. LessRec accepts up to 1 GB per file (about 6 hours of audio).

2. .docx and .srt output

OpenAI returns JSON or plain text. If you want a Word doc with paragraph breaks, you write a converter. If you want SRT subtitles with the right timestamps, you write a converter. LessRec ships .txt, .docx, and .srt in the response.

3. Job queue and retry logic

OpenAI's API call is synchronous — you upload, wait (up to several minutes for long files), get the result. If the connection drops, you re-upload. Building a proper job queue (background worker, retry, idempotency keys, status polling) takes a developer 1-2 days of work. LessRec has all of this and exposes a GET /api/job/:id endpoint.

4. No web UI

Your end users can't drag-and-drop into OpenAI's API directly. You need a web app on top. LessRec is that web app.

5. No CPU-tier pricing

OpenAI runs Whisper on GPUs, charging GPU prices. LessRec runs Whisper large-v3 INT8-quantized on CPUs (faster-whisper) — same model, ~80% of the accuracy at 5-10% of the per-minute compute cost. Our $0.05/min covers all-in compute + storage + bandwidth + UI; OpenAI's $0.006/min is just the GPU API.

When OpenAI's API is the right call

You should use OpenAI's API directly if:

• You're a developer integrating transcription into another product (chatbot, voice notes app, video editor)
• You can write the upload-chunking and job-queue code (~1-2 days)
• You're already paying for OpenAI for other reasons (saves you a vendor)
• You need very low latency (~real-time) and your audio is small (<25 MB chunks)
• You expect 10,000+ minutes per month of volume — at that scale, the $0.044/min savings adds up to $440+/mo, which is more than enough to justify the engineering work

When LessRec is the right call

• You're an end user (podcaster, journalist, lawyer, course creator) who doesn't want to write code
• You transcribe long files (1-6 hours each) and don't want to chunk them by hand
• You need .docx or .srt output without writing converters
• You don't want a $20/mo subscription if you only transcribe a few hours per month
• Volume is <1,000 minutes/month — the engineering time to build your own wrapper would cost more than 2 years of LessRec usage

The hidden cost developers forget

Many developers see $0.006/min and start building. Then they hit:

• Files > 25 MB → need ffmpeg-based chunking with overlap to preserve timestamps across chunks
• Long jobs → need background queue (BullMQ / Sidekiq / Celery), Redis, retry storms
• Customer needs SRT → need to convert OpenAI's segments to SRT (handle timestamp accuracy, formatting)
• Customer needs speaker labels → OpenAI doesn't do diarization at all; need pyannote.audio or alternative, GPU costs explode
• Storage → uploaded files need to live somewhere (S3) until transcription completes; bandwidth costs

By the time you ship a production-grade wrapper, you've spent 1-2 weeks of engineer time (~$5,000-15,000 fully loaded) and you're maintaining infrastructure forever. Whether that's worth it depends on your volume.

What LessRec actually offers (under the hood)

For transparency: we run Whisper large-v3 INT8 (faster-whisper port) on Hetzner CX43 dedicated servers. CPU-only inference at ~50-100× realtime. Job queue is SQLite + a simple Python worker. Storage is local disk with 7-day cleanup. Stripe for billing. Express for the API. The whole stack is ~1,500 lines of code maintained solo. We charge $0.05/min and the marginal cost per minute is ~$0.001, so the margin pays for the engineering time you'd otherwise spend.

FAQ

Do you have an API endpoint or just web upload?

POST /api/transcribe with multipart form-data, GET /api/job/:id for status, GET /api/job/:id/download for output. API keys are issued after first paid pack — email hello@lessrec.com.

Can I self-host Whisper instead?

Absolutely. faster-whisper is open source. On an M-series Mac it runs at 5-10× realtime (free, your machine). On a $40/mo Hetzner CX22 it runs at ~25× realtime. Worth it if you transcribe enough volume to justify the setup time.

Why CPU instead of GPU?

For Whisper large-v3 INT8, a CPU at 50-100× realtime costs $0.001/min in compute. A GPU at 200× realtime costs $0.005/min. Both are "instant" from the user's perspective. We chose CPU because it lets us undercut everyone else and still have margin.