Customer-facing updates to the Anoman gateway, dashboard, docs, and pricing. Reverse chronological.
May 2026
Brand-new authentication, errors, streaming, rate-limits, webhooks, regional-routing, and per-endpoint reference pages. Conceptual deep-dives for weighted tokens, the guardrails pipeline, and caching. Two migration guides (from OpenAI, from Anthropic) and a recipes cookbook covering code-review bots, streaming chat UIs, and batch document analysis.
9 text + 3 vision OSS models added to the catalog with full data-residency disclosure. Up to 75% cheaper than prior US-routed paths for the same weights. Auto-failover when the primary route degrades.
The 03:00 UTC nightly sync now discovers new model launches from OpenAI, Anthropic, Google, DeepSeek, Mistral, Groq, xAI, OpenRouter, Fireworks, and SiliconFlow — plus a pricing refresh from a maintained public pricing dataset. New model launches typically surface within 24 hours. The /models page also refreshes every 5 minutes (was 1 hour).
Model card + browser + detail page now lead with the data-processing region (🇮🇩 Indonesia / 🇸🇬 Singapore / 🇨🇳 China / 🇺🇸 US / 🇪🇺 EU / 🇯🇵 Japan / 🌐 Global) instead of upstream supplier identity. The /pricing page added a Models-by-Region section. Legal pages generalized infrastructure references to 'Indonesia data center'.
Mobile menu wouldn't scroll past the Company section, hiding Docs + Models + Pricing buttons on smaller phones. Added overflow-y-auto + overscroll-contain to the SheetContent. Expanded the Docs nav with 7 entries (Quickstart, Authentication, Streaming, Rate Limits, Errors, Webhooks, API Reference).
Internal ops dashboard fully rebuilt on the @anoman/ui design system. New Billing Admin + Incidents + Usage Analytics pages. Sidebar restructured into 5 grouped sections.
All 14 dashboard pages rebuilt with KPI strips, page headers, data tables, and filter bars from the shared design system. New Feature Explorer page (1-click guardrail enablement) and Audit Log page (SOC 2 evidence with NDJSON export).
Rate limits now enforce per customer instead of per API key. N keys on the same account share one bucket — closes the multi-key gaming exploit. Tier defaults are DB-backed and editable from the ops dashboard. Per-customer overrides supported.
New @anoman/ui workspace: 25 vendored shadcn primitives + 13 Anoman patterns (KPI strip, data table, status pill, sparkline, tier badge, page header, filter bar, sidebar, top bar). Foundation for the marketing / dashboard / ops revamps that followed.
Server-Sent Events streaming on /v1/chat/completions, /anthropic/v1/messages, and the dashboard playground. Sub-second time-to-first-token typical. Mid-stream errors signaled via in-band error frame so HTTP-200 streams still surface upstream failures. Edge proxy timeout no longer bites long completions — each chunk resets it.
April 2026
End-to-end validation across model coverage, batch flow, full guardrail moats, all 7 tier rate limits, response transparency headers, perf under load, SSE streaming, public attacker pose, billing, frontend rendering, data plane consistency, Anthropic adapter, audit events, vision, auth, SDKs. Caught + fixed 2 production bugs in the process.
A worker pool offloads injection and PII inference from the main event loop. Per-key inflight cap (3–100 depending on tier) with structured 429 responses. 9-scenario benchmark suite. Master guardrail toggle on policy groups (audit-logged when used).
Configurable conversational guardrails per policy group. YAML + flow-rule editors in the dashboard. Audit events on config change. Packaged as an optional install extra so the base gateway image stays lean.
Per-customer Isolation Forest models trained nightly from a 14-day analytics lookback, scored inline alongside the existing Z-score rules. 8-dim feature vector (total_tokens, cost, request_rate, tool_call_count, guardrail_rate, session_length, prompt_len, hour_of_day). Fail-soft: exceptions never block the completion pipeline.
Model Context Protocol server registry with per-tool RBAC, OAuth2 scope check, allow/deny/require_approval policy modes. Latency-based provider routing reorders fallbacks best-first based on the 5-min rolling p50 + error rate window.
$9/mo platform fee + raw provider cost with progressive markup (0% under 1K monthly calls, scaling to 5% past 50K). Always cheaper than reseller alternatives at typical volumes. An in-memory counter per customer per month tracks the markup bucket.
JSONB guardrail_overrides on API keys for surgical opt-outs. Every completion response now carries an _anoman metadata block (guardrails / routing / cache / cost / weighted-tokens / burst). Burst credits monetize temporary rate-limit bumps (2x/5x/10x for 1h/4h/24h).
March 2026
Four PII modes per policy group: redact (default), tokenize (reversible, restored in response), synthetic (replace with realistic fake values), block (return 403). 8 entity types: PERSON, EMAIL, PHONE, CREDIT_CARD, SG_NRIC, ID_NIK, IP_ADDRESS, URL.
Append-only audit_events table for SOC 2 evidence. NDJSON streaming export. SAML SSO with a session-auth bridge. Provider-side attack signature pre-screening for enrolled providers.
Anthropic + OpenAI + Google native batch APIs integrated. 50% off realtime cost. Dual-mode batch worker (immediate vs native) with tier-aware routing. Anomaly + SLA breach + provider-down ops email alerts.
Want to be notified of changes? Subscribe to our JSON feed or follow the blog for deeper write-ups.
Major changes always get a deeper write-up on the blog.
We use cookies to improve your experience. Necessary cookies are always active. Learn more