backend new models check - together - 2026-09-16

Files checked: `src/backend/drivers/ai-image/providers/together/models.ts`, `src/backend/drivers/ai-video/providers/together/models.ts`

Note: together **chat** is fetched live from the API and cached (no hardcoded list), so it is not checkable here — only together image and video have `models.ts` files.

**image (4)**
- `ByteDance/Seedream-5.0-lite` | pricing: $0.035 for both 2K & 3K outputs | notes: new model, distinct from backend's Seedream-3.0/4.0
- `openai/gpt-image-2` | pricing: $0.053 (0.006–0.165 by size/quality) | notes: new model, backend only has `gpt-image-1.5`
- `google/flash-image-3.1-lite` | pricing: $0.069/image | notes: new model, distinct from backend's flash-image-3.1 (full) and flash-image-2.5
- `prunaai/p-image-ideogram` | pricing: from $0.00225/image | notes: new model

**video (5)**
- `openai/sora-2` | pricing: $0.80 (720p/8s) | notes: new model
- `openai/sora-2-pro` | pricing: $3.00 (1080p/8s) | notes: new model
- `google/veo-3.1-test-debug` | pricing: $0.08 (description: "test") | notes: id literally contains "test-debug" and description says "test" — likely a non-production/staging entry upstream, not a real model to add; verify intent before treating as new
- `HappyHorse/HappyHorse-1.0-T2V` | pricing: $0.10 per 5s of video | notes: looks like a duplicate upstream listing of the model already present as `alibaba/happyhorse-1.0-t2v` (same model, different org-prefix) — flagging for verification rather than confirmed new
- `MiniMaxAI/MiniMax-H3` | pricing: $0.1391/s of generated output @2k | notes: new model

skipped (non-image/video types): 211 (171 chat, 14 language, 1 code, 1 embedding, 2 rerank, 7 transcribe, 15 audio)

Start with src/backend/drivers/ai-image/providers/together/models.ts and src/backend/drivers/ai-video/providers/together/models.ts, comparing their existing model and pricing entries with Together’s current catalog. Verify whether the test-debug and HappyHorse listings are production models or duplicates before adding them. Done means confirmed image and video models are represented with accurate pricing and questionable entries are resolved.

Written by the indexing model from the issue text.