The NeverNo API

Every request goes to the base URL below and carries your key as a bearer token, or in an x-api-key header as the Anthropic SDKs send it. Keys start with nn_live_.

The same request and response shapes as OpenAI (and as Anthropic for /messages), and Venice’s names for the image tools.

Send model and messages (roles system or developer, user, assistant and tool, up to 2,000 messages). Use auto and NeverNo picks the cheapest model able to answer well, or name any chat model from /models. temperature, top_p, stop and seed are passed to the models that take them and left out for the others (their supported_parameters in /models).

Images can be read by the models with the vision trait: pass a content part of type image_url, with a public https address (answering directly, 4 MB at most) or a data:image/png;base64,… URL (up to 20 images, 4.5 MB per request). Each image must be a still PNG, JPEG or WebP, 256 to 4096 pixels a side; it is checked against the hard limits and the account’s content filter before the model sees it (no one who could be a minor), then sent re-encoded without metadata. The checks are billed with the reply, at the same rate. max_tokens (or max_completion_tokens) caps the reply; reasoning_effort turns on reasoning for the models that support it, returned in reasoning_content.

Web search: enable_web_search set to on searches the web (Google results) before every reply, auto only when the question needs fresh facts, off never (the default). enable_web_scraping: true reads the web pages linked in the last message. The model cites the pages as [1], [2]…, numbered like web_search_citations (title, url, snippet), returned in the response and in the first streamed chunk. Both are billed in credits at their real cost × 2.5, with the reply (about 0.25 credit per search). They can also be passed inside venice_parameters.

Pass tools (functions with a JSON schema), and optionally tool_choice (auto, none, required or one named function) and parallel_tool_calls. When the model calls a tool, the reply holds tool_calls and finish_reason is tool_calls; run the function, then send the assistant message back with a tool message holding the result (tool_call_id), and the model answers from it. The models whose capabilities.tools is true in /models support it, auto included; any other model answers with a clear 400 tools_not_supported.

response_format: { "type": "json_object" } asks for valid JSON, and json_schema for JSON that follows your schema, on the models listed with response_format and json_schema in their capabilities.

POST /messages speaks Anthropic’s format: system, text, image, tool_use and tool_result blocks, tools with input_schema, tool_choice, stop_sequences, thinking, and the streaming events (message_start, content_block_start, text_delta, input_json_delta, content_block_stop, message_delta, message_stop, ping). The key goes in x-api-key or Authorization: Bearer; anthropic-version and anthropic-beta are accepted.

Claude model names (claude-sonnet-…, claude-opus-…, claude-haiku-…, with or without a date) are served by the Claude model of that family listed in /models, and any other chat model id of the catalog works too: deepseek-v4-pro or glm-5.2 through the same endpoint cost far less. The model that answered is named in the X-NeverNo-Model header. Anthropic’s own server tools (its web search, its code execution) are left out; your client’s tools work as usual.

Any tool that accepts an OpenAI-compatible provider works with the base URL and a key; the models with tools in /models handle agent work.

Send a prompt and, optionally, a model (default z-image-turbo). Images are returned in b64_json only: they are never stored on our side, so there is no URL to fetch later. Decode the base64 and save the file yourself.

Every image is about one megapixel: the size you ask for is matched to the closest supported aspect (1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, 21:9), and the size actually produced is returned in the response.

/embeddings returns the vectors of the model behind NeverNo’s memory (qwen3-embedding-0.6b, 1,024 values, up to 128 inputs a request); dimensions shortens them and encoding_format: "base64" is understood. OpenAI’s model names are accepted and served by it.

/audio/speech reads up to 4,096 characters in MP3 with the app’s voices: tts-kokoro (default, many languages), tts-xai, tts-qwen3, tts-inworld and tts-minimax; voice is a voice name listed in /models, and OpenAI’s names (tts-1, alloy, nova…) are mapped to Kokoro. /audio/transcriptions takes a multipart file of up to 1 MB (WebM, Ogg, MP4/M4A, MP3, WAV, AAC), billed on its duration, and answers in json, text or verbose_json.

1 credit = $0.01. API usage is billed at the real cost of the model you call, times 1.5 for open and closed models (Claude, GPT, Gemini, Grok) alike and times 2.5 for images, with no subscription required. It is paid from a separate API balance: a plan’s included credits and credits bought for the app are never spent by the API, and the API never spends them.

Top-ups of $10, $25, $100, $500 or $2,000 (any amount from $10 to $2,000) carry a volume bonus calculated on the amount paid in one payment: +6% from $100, +7% from $500, usable on open models and images. Taxes are included in the price, charged in your local currency.

Starter credit: the first time you create an API key, your account receives $1 once, usable on open models and images only (a daily budget is shared by all new accounts, so it can run out). The “Dev” subscription, $25 a month, adds $1 of API credit each day, for open models and images only, not carried over to the next day; its request limit is the one of Pro+.

Before a call, the most it can cost is reserved from your balance; once it is done, you pay the real amount and the rest comes back. A call that fails on the provider’s side is refunded.

What a call cost is returned, in credits, in the X-NeverNo-Credits header of the non-streamed answers of /chat/completions and /messages (for example X-NeverNo-Credits: 0.8421, rounded to 4 decimals); a streamed answer sends its headers before the end, so it carries none. When a reply is declined, it may be given back within bounds: the header then reads 0 and nothing is deducted.

Claude models: tokens read from Anthropic’s prompt cache cost 0.025 to 0.1 × the input price depending on the model (0.05 × for Opus and Sonnet) and tokens written to it 1.25 × (2 × for a one-hour cache); /models lists cache_read_credits and cache_write_credits. The reservation assumes the worst case (everything written), the settlement uses what the provider reports.

OpenAI models write the whole prompt to their cache on every call and are billed 1.25 × the input price for it: their input_credits in /models already include it; tokens read back cost cache_read_credits.

Each account has a number of API requests per rolling 24 hours, by plan. Past it, calls get a 429 until the window moves.

Errors use OpenAI’s format, so client libraries show the message as is: { "error": { "message", "type", "code" } }.

We do not log or keep your prompts, your messages, the replies or the images. Nothing of a request is used to train a model. Private models run on providers that keep nothing; anonymized models are reached without your identity.

What we do keep is what billing needs: the time of each call, the model and the credits charged, and when each key was last used.

The same hard limits apply to everyone, through the API as in the app: nothing involving minors, nothing about real people without their consent, no real help to seriously harm others.