This changelog records customer-facing changes to the BabySea API, SDK, dashboard, models, and inference provider support. Subscribe to the status page for incident updates.Documentation Index
Fetch the complete documentation index at: https://docs.babysea.ai/llms.txt
Use this file to discover all available pages before exploring further.
Webhook delivery hardening
May 14, 2026Customer webhook delivery is now more reliable and more strict across the v1 generation API, provider callback recovery, dashboard test/replay actions, dead-letter queue retries, and low-balance credit alerts. BabySea now validates webhook destinations before delivery. Endpoints must use public HTTPS on the default HTTPS port, must not include URL credentials, and must not depend on redirects. Delivery rejects localhost, private, link-local, reserved, multicast, and translated IPv6 destinations, including hosts that resolve to blocked addresses. Outbound webhook delivery now uses a single signed delivery path with DNS pinning, redirect-free HTTPS requests, bounded response handling, and explicit wall-clock timeouts. Retry timing is now kept short and predictable so provider callbacks can finish promptly: immediate attempt, then 0.5s, 1.5s, 3s, and 5s retries. Generation terminal events are now deduplicated more aggressively when provider retries overlap, reducing duplicate
generation.completed, generation.failed, and generation.canceled customer webhooks. Disabled endpoints now receive dead-letter queue entries consistently when subscribed events occur, including credits.low_balance alerts.
Fal and Replicate callback validation is also more resilient. Incoming provider webhook bodies are size-limited while streaming, Replicate signatures now only accept v1 signature components, and cached signing metadata is reused during temporary provider metadata refresh failures.
No customer webhook payload or signature format changed. Existing consumers should continue verifying X-BabySea-Signature with the same t=<timestamp>,v1=<hmac> format.
API usage logging reliability
May 13, 2026API key usage logging across all v1 routes is now awaited with explicit error handling. Logging failures are captured and surfaced in server logs without affecting the response returned to the caller.
Playground
May 12, 2026The browser playground now supports the full generation workflow end-to-end. Select any image or video model from the catalog, build a request using the schema-aware input panel, and send it directly against the US, EU, or JP API with your API key. The playground includes a model schema viewer, per-field validation, cURL export for every request, and a response panel with syntax-highlighted JSON. While a generation is in-flight, a live loading timeline shows the active provider, elapsed time, and failover steps as they happen, backed by Supabase Realtime for instant DB-driven status updates rather than client-side simulation. Content management operations: retrieve, cancel, and delete, are available in a separate panel below the generation section.
Predictive provider routing and regional event pipeline
May 11, 2026
generation_provider_order: "fastest" now routes to providers using real-time rankings computed from generation latency and success data across all regions. Rankings are updated continuously and served from a low-latency cache, so fastest selects the best-performing provider for a given model at request time rather than using a static order.
The platform now runs a regional event pipeline across US, EU, and JP. Generation events flow through a transactional outbox into a multi-region Kafka cluster and are ingested into a streaming analytics layer. This powers the predictive routing above and improves the accuracy of provider health signals used for failover.
These changes are fully transparent, no API or SDK changes are required. fastest behavior improves automatically.
Idempotency, concurrency, alerts, and provider order
April 29, 2026Generation requests now support the
Idempotency-Key header. Reusing the same key with the same request body returns the original response for 24 hours and includes Idempotency-Replayed: true. The TypeScript SDK exposes this through client.generate() with options.idempotencyKey and sets idempotency_replayed on replayed responses.
Accounts now have per-plan concurrency caps for in-flight image and video generations. Requests above the cap return BSE1009 with type concurrency_limit_exceeded. The low-balance alert default is now one $0.50 threshold, and existing custom thresholds remain unchanged.
Multi-provider models now accept fastest for generation_provider_order, and fastest is the default when the field is omitted. Explicit provider orders such as replicate, fal still override the default.
See idempotency, concurrency limits, and low-balance alerts.
Error handling and dashboard improvements
April 28, 2026Error classification is more consistent across inference providers. Rate limits, capacity errors, service pause states, and provider-specific failures now map to clearer
BSE codes so failover and retry behavior are easier to reason about.
Credit usage totals now remain accurate across account history. The dashboard also improves provider display names, usage chart layout, and plan budget totals.
Reliability and API consistency
April 27, 2026Generation attempts now use consistent timeout behavior across inference providers. A slow or unavailable provider no longer blocks the full provider order from continuing. Cancellation and cleanup behavior is more reliable for concurrent requests. Content retrieval now enforces account ownership consistently, and protected health endpoints require API key authentication. Model integrations were updated for provider specification changes, including image editing support,
input_fidelity, and schema consistency.
Execution analytics and observability
April 22, 2026The dashboard now shows generation outcomes, request volume, credit movement, webhook delivery, latency, model usage, and provider usage for each account. Use metrics, logs, credits, and activity to review usage and operational history.
Alibaba Cloud
April 17, 2026Alibaba Cloud is now available as an inference provider. You can select it with
generation_provider_order when the chosen model supports it.
OpenAI and Runway, plus 6 new models
April 8, 2026OpenAI and Runway are now available as inference providers. The update also added 6 model identifiers to the catalog. Models added:
- openai/gpt-image-1.5
- runway/gen4-aleph
- runway/gen4-image
- runway/gen4-image-turbo
- runway/gen4-turbo
- runway/gen4.5
Black Forest Labs and the FLUX 2 models
April 2, 2026Black Forest Labs is now available as an inference provider, with support for the FLUX 2 model family. Models added:
- bfl/flux-2-flex
- bfl/flux-2-klein-9b
- bfl/flux-2-max
- bfl/flux-2-pro
Public launch
March 27, 2026BabySea launched with one regional API for image and video generation, signed webhooks, credit billing, API request logs, a browser playground, and the TypeScript SDK. The launch catalog included 74 image and video models across BytePlus, Cloudflare, Fal, and Replicate. Browse the current catalog in models.