Changelog · UltraGPT API

Changelog

2026-10-12Multi-provider routing: models can have several endpoints (listed in GET /v1/models/{model} → endpoints) with automatic failover between them, load balancing and per-endpoint health. provider.order / only / ignore take endpoint labels; provider.sort orders by price, throughput or latency; provider.max_price (USD per 1M you pay) and preferred_max_latency / preferred_min_throughput. x-ultragpt-endpoint header; provider_endpoint, route, client_disconnected and aborted_on_disconnect on generations. X-UltraGPT-Abort-On-Disconnect: true stops a stream when your client leaves and bills only what was produced.
2026-10-11POST /v1/images/edits (multipart, OpenAI SDK images.edit). Zero-completion insurance now follows OpenRouter exactly: a stream that fails after producing output is billed for what it delivered. Partial image batches return and bill the images that succeeded. Files, batches and video lists are scoped to the workspace (personal or organization) of the key. Idempotency-Key replays now work for /v1/videos and take uploaded files into account. Video submissions require balance for clips still rendering.
2026-10-03OpenAPI spec with response schemas and operationIds (also at /v1/openapi.json) and a Postman collection. POST /v1/files returns 200 like OpenAI; GET /v1/files paginates (limit, after). Invalid limit values are rejected with 400. Account-wide rate ceilings across keys, a per-model circuit breaker (503 model_circuit_open), graceful deploys (503 server_restarting, retryable). Organization member keys are capped by the owner's member spend limit. Data export and erasure under Settings > Your data. Status in Statuspage summary.json format.
2026-10-02/v1/videos (async video generation) and /v1/realtime (WebSocket voice). Per-key spend limits that reset daily, weekly, monthly or never (limit / limit_reset) and a tokens-per-minute limit. Model deprecations with Deprecation / Sunset headers and 410 model_retired. New webhooks: credits.added, key.created / rotated / revoked, video.completed / failed, model.deprecated; failed deliveries now retry across restarts and can be redelivered. Opt-in prompt logging, an audit log, email step-up for sensitive actions, status incidents with RSS and email.
2026-10-01Presets (@preset/slug, versioned), response caching (free hits), response-healing and context-compression plugins, :nitro / :floor / :exacto / :thinking variants, zero-completion insurance, /rerank, /moderations, /search, management keys (/keys), /activity, /generations/lookup, /models filters + /models/count, safety_identifier and X-Client-Request-Id tracking.
2026-09-30Anthropic Messages API, Responses API, legacy completions, embeddings, images, speech, transcription, Files & Batch API, web search plugin, provider routing, Idempotency-Key, key rotation, IP / domain restrictions, read-only keys, webhooks, BYOK, organizations, daily spend limits.
2026-09-29Model fallbacks (models[]), /key, /credits, /generation, app attribution, richer /models.
2026-09-28Private preview: /v1/chat/completions and /v1/models.