> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bytespike.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Multimodal

> How image, video, audio, and embedding endpoints sit alongside text under the same key.

ByteSpike treats every modality as a peer of text — same auth header,
same wallet, same headers. The shape of the request is family-specific
(images use `/images/generations`, video uses `/tasks/submit`, etc.) but
the auth and billing surfaces are unified.

## Family overview

| Family                | Sync?                         | Key endpoints                                                                                     |
| --------------------- | ----------------------------- | ------------------------------------------------------------------------------------------------- |
| Text                  | Yes (with optional streaming) | `/v1/messages`, `/v1/chat/completions`, `/v1/responses`, `/v1beta/models/{model}:generateContent` |
| Image (sync)          | Yes (≤30s)                    | `/v1/images/generations`, `/v1/images/edits`                                                      |
| Image / Video (async) | No, async                     | `POST /v1/tasks/submit` → poll `/v1/tasks/query` (or SSE `/v1/tasks/stream/:id`)                  |
| Utility               | Yes (free)                    | `/v1/models`, `/v1/usage`, `/v1/balance`, `/v1/tasks/{query,cancel}`                              |

## Why async for video

Video generation typically takes 30–180 seconds of GPU time. Holding
an HTTP connection that long is fragile (proxies time out, retries
multiply cost). The async pattern is:

1. `POST /v1/tasks/submit` → returns `task_id` + `estimated_credits` + `estimated_seconds` immediately
2. Poll `POST /v1/tasks/query` with `{"task_id": "..."}` (free) on a cadence matched to the ETA, or stream `GET /v1/tasks/stream/{task_id}` over SSE for push delivery, or register a `callback_url` on submit and skip polling entirely
3. When `status == "completed"`, the `output` array carries the result URLs

Status transitions: `pending → running → completed` (happy path),
`pending → running → failed` (unbilled), or
`pending|running → cancelled` (manual cancel, billing depends on the
model's refund policy).

## Mixing modalities in one app

Each API key is bound to **one** routing group on ByteSpike,
which determines which models it can reach. For an app spanning
multiple model families, create one key per group:

```bash theme={null}
# Key bound to the Claude group — reaches Claude + (via translation) DeepSeek-Anthropic + cross-protocol Gemini-via-translated
curl https://llm.bytespike.ai/api/v1/keys -H "x-api-key: $ADMIN_KEY" -d '{
  "name": "vision-claude-prod",
  "group_id": 1
}'

# Key bound to the Gemini group — reaches Gemini native + image/video models in that group
curl https://llm.bytespike.ai/api/v1/keys -H "x-api-key: $ADMIN_KEY" -d '{
  "name": "vision-gemini-prod",
  "group_id": 3
}'
```

The list of available groups for your account is at
[`GET /api/v1/groups/available`](https://console.bytespike.ai/keys)
or the **Create key** dialog in the console. To attribute spend per
pipeline, hit [`GET /api/v1/usage`](/api-reference/account/me-usage)
filtered by `api_key_id`.