Codex multi-region rotation
Codex inference now cycles across three Azure regions per request group. Failover latency under 80 ms p99. Stale-region eviction triggers on 2 consecutive 5xx responses within a 30-second window.
9-region Azure routing
Expanded the routing mesh to nine Azure regions including Switzerland North and UAE North. Regional preference is now sticky per workspace session token, with cross-region fallback when capacity drops below 15%.
Dropped fake Claude SDK 401
Removed the synthetic 401 challenge that was injected for unverified SDK fingerprints. All legitimate SDK builds now pass the handshake without a preflight penalty. Latency improvement of ~120 ms on cold starts.
27-model catalog
Added GPT-5.1, Gemini 2.5 Pro, Claude Opus 4, Llama 4 400B, and 23 other models to the unified catalog. Each model exposes a consistent /v1/chat/completions surface with provider-specific quirks normalized at the edge.
Streaming token budget enforcement
Per-key token budgets now apply mid-stream. When a key exhausts its quota during a streaming response, the connection receives a clean finish_reason: 'budget' rather than a hard RST. Clients see the partial response up to the cutoff.
WebSocket keepalive rewrite
Replaced the legacy ping/pong loop with a single-frame opcode-9 keepalive that respects Azure's 300-second idle timeout. Memory pressure on long-lived connections dropped by 40%.
Rate-limit headers standardization
All responses now include x-ratelimit-limit, x-ratelimit-remaining, x-ratelimit-reset, and retry-after in RFC 7231 format. Keys provisioned before 2026-05-01 receive a one-time migration header for 72 hours.
Prompt caching GA
Prompt caching graduated from beta. Cache hits are billed at 10% of the base token rate. Cache keys are scoped to (model, workspace, prefix-length). TTL is 5 minutes with LRU eviction under memory pressure.
Audit log export
Workspace owners can now export audit logs as newline-delimited JSON to Azure Blob Storage or S3-compatible buckets. Export jobs respect the same RBAC scoping as the live log viewer.
Fine-tune checkpoint resume
Interrupted fine-tune jobs now resume from the last saved checkpoint instead of restarting. Checkpoint interval is configurable between 50 and 500 steps. Storage cost for checkpoints is included in the job price.
OAuth device flow
Added RFC 8628 device authorization grant for headless environments. The flow presents a user code and verification URI with a 15-minute expiry. Refresh tokens rotate on each use.