Models and pricing
The whole catalogue, with prices in plain sight.
Chat and vision, document OCR, decisions, embeddings, rerank and redaction through one API. Every model with its price per million tokens, read live from the API. No minimums, no lock-in.

Real responses from every route, with their requests and tokens.See real examples
Chat and vision models
They generate text from text and images, in OpenAI’s chat completions format. Input and output are billed separately, and input served from your API key’s prompt cache at the cached price of the table: see how a request is billed, below.
RoutePOST /v1/chat/completions
The catalogue could not be read from the API right now. Try again in a few minutes or check the API response directly.
Open the API response| Model | Context | Input | Output | Cached | Capabilities |
|---|---|---|---|---|---|
Orion Pro Orionorion-proThe top tier of the Orion family, for demanding tasks. | 262,144 tokensMax output: 32,768 tokens | 0.144 | 1.50 | 0.036 |
|
Orion Plus Orionorion-plusThe efficient tier of the Orion family: economical for large volumes. | 262,144 tokensMax output: 32,768 tokens | 0.101 | 0.60 | 0.025 |
|
Qwen3.8 27BOpenqwen3.8-27bDense open-weight model under its usual identifier, quantised to 4 bits for serving. | 262,144 tokensMax output: 32,768 tokens | 0.144 | 1.50 | 0.036 |
|
Qwen3.6 35B A3BOpenqwen3.6-35b-a3bMixture-of-experts (MoE) open-weight model, quantised to 4 bits for serving and built for large volumes. | 262,144 tokensMax output: 32,768 tokens | 0.101 | 0.60 | 0.025 |
|
PaddleOCR-VL 1.6OpenDocument OCRpaddleocr-vl-1.6Open-weight document OCR model: from one page image and a task prompt, it returns the page’s text, tables, formulas or chart data. | 16,384 tokensMax output: 8,192 tokens | 0.016 | 0.016 | – |
|
USD per million tokens
Prices and models read from the API on 20 Nov 2026, 14:37 UTC, when this page was published.
OrionOpenRead the models guideSee a real example
Try it in the playground
Decision models
PreviewDecision models do not generate text: they answer questions about an input with a probability for every possible answer, ready for your code to act on. Use them to route, classify, filter and triage. They are in public preview: any change before general availability will be announced in the documentation.
- Typed answers
- Yes or no, one of several options or a score on a scale, with the probability of every answer. Many questions fit in a single call.
- Two routes
- POST /v1/decisions, in the format of OpenAI’s Decisions API, and POST /v1/systemone, compatible with Jev.
- Only input is billed
- Input tokens are billed at the price in the table. Output is not billed.
The catalogue could not be read from the API right now. Try again in a few minutes or check the API response directly.
Open the API response| Model | Context | Input | Routes |
|---|---|---|---|
Orion Compass Orionorion-compassThe Orion decision model for questions that need depth: nuanced inputs, fine-grained options and detailed rubrics. | 32,000 tokens | 0.034 |
|
Orion Compass Flash OrionImage inputorion-compass-flashThe light version of Orion Compass, at a lower price, for high-volume routing, filtering and triage. It also reads images, such as a scanned document or a screenshot. | 32,000 tokens | 0.017 |
|
d3OpenImage inputd3Open-weight decision model for text and images, under the Apache-2.0 licence and served with its published weights. | 24,576 tokens | 0.042 |
|
USD per million input tokens. Output is not billed.
OrionOpenImage inputRead the decision models guideSee a real example
Try it in the playground
Embeddings, rerank and redaction
Building blocks for search, retrieval-augmented generation (RAG) and privacy, each with its own route and model. Only input tokens are billed.
The catalogue could not be read from the API right now. Try again in a few minutes or check the API response directly.
Open the API response| Model | What it does | Input |
|---|---|---|
EmbeddingsPOST /v1/embeddings | ||
Qwen3 Embedding 0.6BOpenqwen3-embedding-0.6b | Turns text into multilingual 1024-dimension vectors for semantic search, clustering and RAG, in OpenAI’s embeddings format. | 0.008 |
RerankPOST /v1/rerank | ||
Qwen3 Reranker 0.6BOpenqwen3-reranker-0.6b | Scores how relevant each document is to a query so the best results come first, in the rerank format of Cohere and Jina. | 0.008Tokens estimated from characters (characters / 4) |
RedactionPOST /v1/redact | ||
Privacy FilterOpenprivacy-filter | Finds and masks personal data in text: names, email addresses, phone numbers, account numbers and more. | 0.009Tokens estimated from characters (characters / 4) |
USD per million input tokens. Output is not billed.
Prices and models read from the API on 20 Nov 2026, 14:37 UTC, when this page was published.
OrionOpenRead the embeddings, rerank and redaction guideSee real embeddings, rerank and redaction examples
Try it in the playground
Sandboxes
Private previewIsolated microVMs for AI agents, compatible with the E2B SDKs, in private preview. Reserved vCPU and memory are billed per second while a sandbox runs, and what a paused sandbox stores is billed per GiB.
| What is billed | When | Price (USD) |
|---|---|---|
| Reserved vCPU | Running | 0.0504per vCPU-hour |
| Reserved memory | Running | 0.0162per GiB-hour |
| Storage | Paused | 0.08per GiB-month |
| Storage of your own templates, on request | Saved template | 0.08per GiB-month |
- Billed per second, with no minimum: a sandbox that runs for three seconds is billed for three seconds.
- For scale, a sandbox of the default template costs 0.1332 USD for each hour it runs.
- There is no welcome credit for Sandboxes.
- Your own templates are enabled on request, per account. What a template stores, its memory and the changes to its disk, is billed at the price of a paused sandbox until you delete it.
Try it in the playground

Orion: Serenity’s own family
Orion is Serenity’s own model family, built for production: Orion Pro for demanding tasks, Orion Plus for large volumes and, in preview, the Orion Compass decision models. The same API as the open models; the context, capabilities and price of each model are in the tables on this page.
Open models: published weights, usual identifiers
Served from their published weights under their usual identifiers. The open chat models are quantised to 4 bits, as the API’s model list shows.
How billing works
Per million tokens, with no minimum fee and no lock-in. Chat models bill input and output; decision, embeddings, rerank and redaction models bill input only.
How one request is billed
An illustrative chat request: the second turn of a conversation, sent with the same API key as the first, at the prices of the chat table.
orion-plusThe 10,500 tokens of the request
- Cached input: 8,192
- Input: 1,808
- Output: 500
The usage field of the response
"usage": { "prompt_tokens": 10000, "prompt_tokens_details": { "cached_tokens": 8192 }, "completion_tokens": 500, "total_tokens": 10500}| Line | Tokens | USD per million | USD |
|---|---|---|---|
| Input, not cached | 1,808 | 0.101 | 0.0001826 |
| Cached input | 8,192 | 0.025 | 0.0002048 |
| Output | 500 | 0.60 | 0.0003000 |
| Total | 0.0006874 | ||
Without the cache, the same request would cost 0.0013100 USD. 1,000 requests like this one: 0.69 USD.
The prompt cache, one per API key
When a request starts with the same tokens as an earlier request made with the same API key (a long system prompt, a shared document, the earlier turns of a conversation), that prefix can be served from the key’s prompt cache and is billed at the cached input price. Each key has a cache of its own, never shared with another key, so a new key starts with an empty cache. Short prefixes are not cached: keep the stable part at the start of the prompt.
On orion-plus: 0.025 USD per million cached input tokens, against 0.101 USD at the input price.
What counts as a token?
A token is the unit into which the model splits text: a short word, part of a long word or a punctuation mark. As a reference, 1,000 tokens are about 750 words in English and slightly fewer in Spanish. Chat models bill input tokens (your prompt, including the system message and history) and output tokens (the completion), and every response includes the usage field with the exact count. For rerank and redaction, tokens are estimated from characters.
Usage limits
Default limits are generous and are raised under contract. If you exceed them, or the API cannot take the request at that moment, it responds with status 429: retry with an increasing delay.
Monthly cost example with the chat models
Two indicative volumes, calculated with the input and output list prices of the chat table.
| Model | Prototype20M input tokens and 5M output tokens | Production500M input tokens and 100M output tokens |
|---|---|---|
Orion Proorion-pro | 10.38 | 222.00 |
Orion Plusorion-plus | 5.02 | 110.50 |
Qwen3.8 27Bqwen3.8-27b | 10.38 | 222.00 |
Qwen3.6 35B A3Bqwen3.6-35b-a3b | 5.02 | 110.50 |
PaddleOCR-VL 1.6paddleocr-vl-1.6 | 0.40 | 9.60 |
USD per month. List prices, without cache or taxes. How cached input is billed is shown above, in how a request is billed.
Dedicated capacity or an SLA?
For high volumes, specific terms or a service level agreement, we work under contract.
