Whatever you want.
One key, every leading model, lower cost. Built for Codex, Claude Code and your agents.
Get API KeyPricing
Competitively priced model channels, quoted by model and volume. Two more ways to spend less:
Document sorting, asset generation and product tagging can trade a longer turnaround for a lower price.
With your own GPUs or steady, high volume, we work out whether open-source models in your environment cost less.
$1.32Official
$0.792OceanToken · per 1M tokens
$0.375Official
$0.225OceanToken · per 1M tokens
$2.90Official
$1.89OceanToken · per 1M tokens
$4.00Official
$3.00OceanToken · per 1M tokens
Data & plans
Start self-serve, agree enterprise rates, or have our engineers run it inside your environment.
Where can this workload’s data go?
External models, with safeguards
Sign up and pay per use. Every model’s price is in the catalog.
Get API KeyManaged hosting
Lower rates for committed usage, agreed around your volume and service needs.
In your environment
Our engineers deploy and tune it with you.
AI Workspace
Deep Research, marketing and sales assets, and chat that picks up where you left off, in one workspace your company manages.
FAQ
What OceanToken is, what it costs, which models you can call and how it fits your agents. Each answer stands on its own.
OceanToken is one OpenAI-compatible API and one prepaid balance for the frontier open models (DeepSeek, Qwen, GLM, Kimi, MiniMax) at lower prices, alongside GPT, Claude and Gemini, plus image, video, speech, transcription, embedding and rerank models. You create one account, add credits, create an API key, and call any model with the same base URL and the same key. Every model's price, context window and capabilities are listed in the catalog before you call it.
Yes. The OceanToken API speaks the OpenAI API, so the official OpenAI SDKs for Python, TypeScript and other languages work after changing two values: the base URL to https://api.oceantoken.ai and the API key to your OceanToken key. Claude models also answer on the Anthropic Messages API at /v1/messages, so the Anthropic SDK works unchanged with the same two changes. Chat completions, the Responses API, images, video, embeddings, speech, transcription and rerank all follow the same shapes.
The catalog spans 7 types of model: language models (DeepSeek, Qwen, GLM, Kimi, GPT, Claude, Gemini, Grok, Mistral, Llama and more), image generation and editing models, video generation models, text-to-speech, speech-to-text, text embeddings and rerankers. The Models page lists every one with its exact model ID, price per unit, context window and whether it supports reasoning, tool calling or image input. The list is rebuilt from the live service every time the documentation is published.
You pay per use at the price shown for each model, in US dollars, from a prepaid balance; there is no subscription and no platform fee on top of the model price. Language models are priced per 1 million input and output tokens, images per image, video per second, speech per 1,000 characters and transcription per minute of audio. Some models are tagged :free and cost nothing with limited throughput. Each model page shows the current price before you make a call.
Run one command in a terminal and sign in: npx -y @oceantoken/cli@latest connect codex (or claude-code, cursor, gemini, openclaw). It adds the OceanToken MCP server at https://mcp.oceantoken.ai/mcp and 3 skills to that agent, which can then search models, estimate costs and generate images, video and speech from your balance. An AI agent can also follow the instructions at https://mcp.oceantoken.ai/install on its own. The same API key works in any OpenAI-compatible tool by setting the base URL.
Yes. Image models answer on /v1/images/generations and /v1/images/edits, video models on /v1/videos as asynchronous jobs you poll, text-to-speech on /v1/audio/speech, transcription on /v1/audio/transcriptions, embeddings on /v1/embeddings and rerankers on /v1/rerank. All 7 types use the same key and balance, and each request appears in the console's Logs with its cost and latency. Image and video download URLs stay valid for 24 hours, and video jobs accept an Idempotency-Key so a retry can never start and bill a second job.
Yes. Each API key can carry its own credit limit, optionally resetting daily, weekly or monthly, plus an expiry date, so a key for a test project or a teammate cannot draw more than you allow from the shared balance; a key that has used its limit gets a 400 key_budget_exceeded error while the account may still have credits. The console shows every key's usage, and the Logs page lists each request with its model, tokens, cost and latency. Keys are shown once at creation and can be rotated with a 24-hour overlap.
You add credits by card in the console, in top-ups between $10 and $100,000, and every request deducts its exact cost from that prepaid balance at the price listed for the model when the call is made, never retroactively. Purchased credits are valid for 12 months from purchase, and there is no monthly minimum. A request that would take the balance below zero is refused before it runs, and long jobs such as video hold their estimated cost until they finish. Spend shows per day and per model on the Activity page.
Models whose ID ends in :free cost nothing to call, with limited and unguaranteed throughput; the Models page filters show them with a Free label. Paid models require a positive prepaid balance. This lets you test the API shape, SDK setup and agent integration at zero cost, then switch to a paid model by changing only the model ID.
Rate limits are per API key, as requests per minute and tokens per minute, and every response reports them in x-ratelimit-limit and x-ratelimit-remaining headers; exceeding one returns 429 with a Retry-After header. A synchronous request may run for at most 240 seconds and a request body may be at most 10 MB, so long text generations should use stream: true and large files should be sent as URLs. Video generation runs as a job that takes minutes, so you poll /v1/videos/{video_id} instead of holding a connection open. Details are on the Limits page.
More in the documentation FAQ, with every model's price on the models page.