Simple, Transparent Pricing

Every request is billed dynamically, with the metric chosen to match each task: resolution × steps for images, characters for speech, tokens for embeddings, duration with optional resolution for video, hours for transcription, input resolution for OCR, and per-image rates for background removal and upscaling. There are no subscriptions, no monthly minimums, and no hidden fees — you fund a prepaid balance and each successful inference deducts its exact cost. Before any job, you can call the matching /price endpoint to preview the precise cost for the model and parameters you plan to use.

Yes. Every new account receives $5 in free credits the moment you sign up — no credit card, no upfront commitment. That balance is enough to test most models extensively, generate hundreds of images, transcribe several hours of video, or prototype a complete AI pipeline. Your account starts on the Basic tier with conservative rate limits designed for testing; making any payment upgrades you to Premium with no waiting period.

Inference is routed through a globally distributed GPU network rather than concentrated in a few hyperscale data centers, which removes most of the infrastructure markup baked into traditional cloud pricing. We also serve highly optimized open-source models — many of them quantized (INT8, FP8, NF4) and distilled — so each request uses fewer GPU seconds without sacrificing output quality. The combined effect can deliver up to 20× lower inference cost for comparable workloads.

deAPI uses Stripe for secure card payments and supported local methods. You can either make one-off top-ups (with preset amounts of $10, $25, or $50) or enable automatic top-ups, which recharge your balance whenever it drops below $2 — perfect for production workloads that can't afford to fail mid-job. For B2B customers needing custom invoices, larger commitments, or tailored billing terms, our team arranges individual agreements; just reach out to support.

The Basic tier offers conservative rate limits (typically 1–10 RPM depending on the endpoint) ideal for testing. The moment you make any payment via Stripe, your account upgrades to Premium with 300 RPM across all endpoints and unlimited daily requests — instantly, with no application process. For high-volume production needs beyond Premium, dedicated capacity, or enterprise terms, our team can set up bespoke arrangements with volume discounts.

Embeddings are billed strictly by the number of tokens you send in — there is no per-request fee, no minimum, and no charge for the returned vectors. In practice the cost is so low it disappears into rounding: a short search query of about 10 words runs around $0.000002, an email-sized paragraph (~100 words) costs roughly $0.000016, and a full ~1,500-word blog article comes in at about $0.000279. That means you can embed an entire knowledge base of tens of thousands of documents for just a few dollars, which makes large-scale RAG and semantic search projects financially viable from day one.

The endpoint is powered by BGE M3, a modern open-source embedding model designed for high-quality semantic representation across a wide range of content types and languages. For exact technical specifications such as vector dimensionality and maximum input length, the Model Selection endpoint exposes the canonical metadata at request time — so your client always sees the authoritative spec, even if the model is updated. In practice this gives you production-grade embeddings without having to host or fine-tune anything yourself.

You can batch up to 2,048 input texts per request, with each individual text capped at 8,192 tokens and a hard ceiling of 300,000 tokens total per request. Batching is the single biggest performance win available on this endpoint: instead of firing thousands of sequential calls when indexing a corpus, you group texts into batches close to those limits and dramatically cut both wall-clock time and rate-limit pressure. For very large datasets, simply chunk your corpus into batches that respect the 300k-token total and process them in parallel.

Embeddings turn text into numerical vectors where semantic similarity becomes geometric distance — so two sentences that mean the same thing end up close together in vector space, even if they share no words. The endpoint is built for five core use cases: semantic text search, document similarity matching, retrieval-augmented generation (RAG) for AI assistants grounded in your own content, text classification and clustering, and recommendation systems. If your application needs to find, group, or compare meaning at scale, this is the building block.

Yes — embeddings generated through the API can be used freely in commercial products, internal tools and customer-facing applications, with no per-seat or per-query licensing on the output vectors. The vectors you receive can be persisted in any vector database (pgvector, Pinecone, Qdrant, Weaviate, Milvus, Chroma, Redis, Elasticsearch, and so on) or even in a flat file for smaller projects. Because the vectors themselves are just numbers, you keep full ownership of your index and can rebuild, migrate, or query it without ever calling the API again — you only pay when you generate new embeddings for fresh content.