What changed, and when.

Added `stream_options.include_usage` field to Chat Completions. Token counts now stream in the final chunk.

Usage graphs now render at day-granularity for date ranges over 7 days. Previous behavior was always hour-granularity.

Invoice PDFs now include line-item model breakdown. Previously invoices showed a single total.

DeepSeek-V3 available in all three regions: us, eu, in. Previously us-only beta.

JSON mode now enforced server-side with schema validation. Malformed JSON responses retried automatically up to 2 times before returning an error.

Dedicated Deployment drain now completes in under 60 seconds for B200 configurations. Previous drain times were up to 5 minutes.

Customer-managed encryption keys (CMEK) now available on Dedicated Deployments in us and eu regions. AWS KMS and GCP KMS supported.

LoRA adapter export now supports GGUF format in addition to HuggingFace safetensors. Download from the console after training completes.

Audit log now filterable by actor (API key, user, system). Exported CSV includes actor field.

Rate limit headers now included on all responses: `X-RateLimit-Limit`, `X-RateLimit-Remaining`, `X-RateLimit-Reset`.