Image, audio, embeddings & fine-tuning pricing
Text isn't the only thing AI APIs charge for. Images, speech, embeddings, fine-tuning and video each use their own pricing unit, per image, per minute, per generation, or per training token. Here are representative 2026 ranges, with a link to verify each on the provider's official page.
Image generation
Billed per image (varies by resolution & quality), or per output token for token-based image models.
Frontier image models from the big labs sit in the few-cents-per-image range, while open models (Flux, Stable Diffusion) served by GPU marketplaces can be much cheaper at lower resolution.
| Provider / model | Typical price |
|---|---|
| OpenAI GPT Image / DALL·E | ~$0.04–$0.12 / image |
| Google Imagen / Gemini image | ~$0.039 / 1024×1024 |
| Flux / Stable Diffusion (Replicate, fal.ai) | from ~$0.002 / image |
Audio — speech-to-text & text-to-speech
Usually billed per minute of audio (transcription) or per token/character (speech synthesis).
Transcription is one of the cheapest AI services; speech synthesis is a little more, and real-time voice models cost more again.
| Service | Typical price |
|---|---|
| Speech-to-text (Whisper-class) | ~$0.006 / minute |
| Text-to-speech (Gemini TTS) | ~$0.037 / minute |
| Real-time / live voice | higher, see official |
Embeddings
Billed per token, like text, but at a tiny fraction of chat-model rates.
Embeddings turn text into vectors for search, classification and retrieval (RAG). They are dramatically cheaper than LLM inference, typically a fraction of a cent per thousand short queries, which is why they're the first tool for cutting cost in search and retrieval pipelines.
Fine-tuning
Training billed per token (sometimes plus an hourly fee); fine-tuned inference is usually priced above the base model.
Fine-tuning adapts a model to your data. You pay once to train, then a (usually higher) per-token rate to run the customized model. As a rough mid-2026 reference, training a mid-size model runs on the order of a few dollars per million training tokens, with inference somewhat above base rates.
| Stage (example) | Rough rate |
|---|---|
| Training (mid-size model) | ~$0.80–$3 / 1M tokens |
| Fine-tuned inference | above base model |
Note: availability is changing, some providers are scaling back self-serve fine-tuning. Confirm on the official page: OpenAI · Google
Video generation
Billed per generation (clip), which depends on length and resolution.
Video is the most compute-heavy category, so it's the most expensive, a single clip can cost from a few cents to a few dollars depending on length and quality.
| Category | Typical price |
|---|---|
| Video generation (per clip) | ~$0.13–$3.20 |
Verify: Google (Veo) · Replicate · fal.ai
All figures are approximate ranges observed in mid-2026 from official provider documentation, shown for orientation only, not live data and not exact quotes. Prices change frequently and depend on model, resolution, quality and region. Always confirm on the provider's official pricing page before budgeting. For text-token models, use the live TokenSwarm calculator. Not affiliated with any provider.