BUILD WITH EVA
Your developer workspace.
One API. Your applications. Usage you can see and control.
CLIENT DOCUMENTATION · OCTOBER 10, 2026
Download the complete API specification.
Account setup, API keys, text and image generation, durable queues, live video jobs, Stripe credits, pricing comparisons, request examples, errors and limits.
Download specification (PDF) · Editable specification (Word)
Pricing overview: PDF · Word. No account required to download.
DURABLE GENERATION QUEUE
Submit once. Pick up your result.
Text and image requests enter separate queues, with one active request per service and fair turns between accounts. Accepted work survives a server restart. Up to 100 text jobs and 50 image jobs can be outstanding, with a combined limit of 10 per account. Existing rate and credit limits still apply.
Add Prefer: respond-async to return a job immediately. Without it, the API waits up to 15 seconds: it returns the normal 200 result if ready, or a 202 job response if work is still pending. Always handle both outcomes.
POST /v1/generate
Authorization: Bearer YOUR_API_KEY
Idempotency-Key: unique-request-id
Prefer: respond-async
Content-Type: application/json
{"prompt":"Explain gravity simply.","max_new":128}
GET /v1/requests/{id} # Poll status
GET /v1/requests/{id}/result # Retrieve completed text or PNG JSON
POST /v1/requests/{id}/cancel # Cancel waiting workUse the same bearer key for all calls. Poll every 2–5 seconds. States are queued, running, completed, failed, cancelled and indeterminate. Results retain the original text/image response format. A result read while pending returns 202; a terminal unsuccessful result returns 409; expired results return 410.
Waiting jobs expire after 15 minutes; a job started before that deadline may finish later. Prompts are encrypted while queued or running, then removed from the live queue record. Completed results are encrypted and retained for 24 hours. Save outputs in your own application. The accounting ledger keeps metadata. These retention periods do not shorten backup retention.
Credit is reserved on acceptance and charged only on success. Cancellation and known failure release it. Uncertain outcomes retain the reservation and block their lane for operator review. Known busy rejections are retried automatically; uncertain completions are never blindly rerun. Reusing the same idempotency key retrieves the existing queued job or retained result. A different body with that key returns 409. A full queue returns 429 with Retry-After.
Queue reads and cancellations share a 120/minute/account limit. Video continues to use its separate video-job endpoints and one-active-job-per-account limit.
USAGE PRICING · USD
Half the selected OpenAI benchmark rates.
Text: $1 / 1M input tokens and $5 / 1M output tokens. Images: $0.0265 each. Video: $0.05 per second.
Text benchmarks GPT-6.1 Sol standard short-context rates. Images benchmark GPT Image 2 medium 1024×1024 output; image prompts are included. Video benchmarks historical Sora 2 720p pricing, retired September 24, 2026. Eva uses different models and output specifications; this is a unit-price comparison, not a quality guarantee.
Rates fixed as of October 5, 2026. Text billing counts submitted prompt and returned text with the o200k_base tokenizer. Output billing is capped at max_new tokens. If the model reports prompt truncation, input is not charged. No cache discount or separate reasoning-token charge applies. We reserve prompt cost plus the output cap before generating, then release the unused amount. Video reserves $0.05 × requested seconds and charges that amount only on success. Known failures and cancellations release the reservation.
$10 in welcome credit. Top up $10, $25, or $50. Credits do not expire; request and service limits still apply. Older completed requests keep their original charge. OpenAI text rates · Image benchmark · Historical video rates
API REFERENCE
Make your first request.
Create an account and API key, then send a prompt. Your first $10 of usage is free. Every request needs a unique idempotency key; keep the same key when retrying an uncertain network request.
curl "https://evaxapi.com/v1/generate" \
-H "Authorization: Bearer $EVA_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{"prompt":"Explain recursion simply.","max_new":128}'Request & response
POST /developers/v1/generate
| Field | Accepted values |
|---|---|
| prompt | Required text · up to 24,000 UTF-8 bytes |
| max_new | 1–8,192 output tokens · default 256 |
| temperature | 0–1.3 · default 0 |
{
"id": "request-uuid",
"object": "generation",
"model": "eva-model-version",
"text": "Eva’s response",
"usage": {
"input_tokens": null,
"output_tokens": 42,
"billable_requests": 0,
"free_credit_used_cents": 0.022,
"cost_usd": 0.00022,
"pricing_version": "openai-half-2026-10-05",
"billing_input_tokens": 10,
"billing_output_tokens": 42
}
}Input-token counts are unavailable from the current model service. The billing ledger stores metadata; queued prompts and results use temporary encrypted storage. Text and image creation have separate endpoints. Streaming and tool calls are not exposed by this API. Video creation uses background jobs described below.
Create an image
POST /developers/v1/images/generations (alias: /developers/v1/images/generate)
Use the same bearer token and a new Idempotency-Key. Send a prompt of up to 1,000 characters. Optional fields: seed (0–2,147,483,647, default 901), steps (10–100, default 30), and guidance (1–10, default 7). The model may use its fixed sampling settings.
{
"prompt": "A red sailboat on a calm blue lake",
"seed": 901
}The response contains one PNG in data[0].b64_json, with width, height and seed. Decode the base64 to save the image. Each successful image costs $0.0265; text, images, and videos share your free and purchased credit balance.
Queued image results are available for 24 hours through the request result endpoint. Save each response in your application.
Create a video
Send POST /developers/v1/videos/generations with your bearer token, a unique Idempotency-Key, and JSON containing prompt (1–1,000 characters) and optional seconds (3, 5, 10, or 15; default 3). A 202 response contains the job ID and status URL. Poll that URL every 15–30 seconds. Generation runs asynchronously; completion time depends on clip length and queue load.
Completed jobs include a download URL; fetch it with the same bearer token to save the MP4. Clips are silent, 832×480 at 16 fps, and retained by the private video worker for 30 days. POST /developers/v1/videos/ID/cancel to cancel and DELETE /developers/v1/videos/ID to remove a finished clip. $0.05 per requested second is reserved per job and consumed only on success; failed or cancelled jobs release it. Deleting a completed video does not refund credit. Video prompts and files are stored on the worker until deletion or expiry. Call GET /developers/v1/capabilities with your bearer token to check which generation types are enabled before submitting work.
Retries, limits & errors
- 401: Missing, invalid or revoked API key.
- 402: Your available credits are exhausted or reserved. Add credits in Billing & payments.
- 409: The idempotency key conflicts with a different request, a result is not successful, or a video was already recorded. Queued text/image retries reuse the original job.
- 429: Account rate limit (30 attempts/minute), monthly limit or model capacity. Retry after a short delay.
- 502/503: Model unavailable, uncertain completion, or API not enabled. Uncertain requests require operator review and are not billed automatically.
Text and images use separate durable queues with one active request per service. Video permits one active job per account and is subject to the worker’s queue limits. Do not use a new idempotency key merely to retry a timeout.
GET /developers/v1/models lists the enabled model and requires the same bearer token.