New
Project Presets & Playground System Prompt Support
- •Project presets — save API call templates (model, system prompt, tools, max tokens) in projects and launch them in the Playground
- •Preset CRUD: create, edit, delete presets with inline editing in the project dashboard
- •Playground now reads ?system= URL parameter to pre-fill system prompts from presets
- •API key selection in Playground auto-sets the key's default model or nest
Fixed
Cloudflare Error Detection Fix
- •Fixed error normalizer misclassifying provider balance/auth errors as "Cloudflare WAF blocks" for all Cloudflare-fronted providers (OpenAI, Anthropic, Groq, Mistral, etc.)
- •New "Insufficient balance" error classification for 402/403 responses with balance/quota messages
- •Rate limiter now returns proper 429 status instead of 500
- •Auth and intelligence endpoints excluded from global rate limiting
Improved
Playground Responsive UI & Rankings Fix
- •Playground header buttons (Fusion, Inspector, Session, Clear) collapse into overflow menu on small screens
- •Rankings tiebreaker — models with equal ranking values sorted by intelligence score
- •Models page converted to SSR with pagination (100 per page) and in-memory caching
Models
10 New Providers Activated
- •Z.ai (GLM), Anyscale, Baseten, Clarifai, Inflection AI, Lambda Labs, OVHcloud, RunPod, Scaleway, Upstage — now available for routing
- •Deploy process improved to prevent accidental image tag rollbacks
New
Image & Video Generation
- •ant:image-gen and ant:video-gen virtual models for media generation routing
- •Image/video generation buttons in the Playground UI
- •Supports Gemini Flash Image, GPT-5 Image, and Flux models
- •OCR bridge: ant:ocr virtual model bridges chat completions to Mistral OCR API
New
Free Model Scanner
- •Automated scanner checks all providers for free models every 6 hours
- •Discoveries stored in free_model_discoveries table for admin review
- •Admin endpoints: /api/admin/free-models/discoveries, /scan, /scan/status
Models
9 New Provider Connectors
- •Perplexity, SambaNova, AI21, MiniMax, Alibaba (Qwen), Moonshot, Venice, Featherless, Chutes
- •Full connectors with fetchModels(), catalog sync, rate limiters, and admin routes
Improved
Topology Redesign
- •3-tab system: Live Flow, Route Replay, System Map
- •Real-time request flow visualization with animated edges
- •Detailed trace inspection with timing breakdowns per stage
New
Reasoning Orchestration
- •Cross-provider reasoning system with reasoning_effort parameter wiring
- •Virtual pools: ant:reasoning, ant:reasoning-fast, ant:reasoning-premium
- •Graduated scoring in router for thinking-capable models
- •DB: reasoning_intensity + reasoning_effort_sent columns in request_traces
Models
First-Class Provider Connectors
- •OpenAI, DeepInfra, Together AI, DeepSeek added as first-class providers with dedicated connectors
- •Each has: fetchModels(), known-model enrichment, catalog sync admin route, direct routing bypass
- •Dual-mode providers: Mistral, NetMind, OpenRouter support both free pool and paid credits
Improved
Classification Pipeline Rewrite
- •New multi-stage classification pipeline with heuristic pre-processing and LLM escalation
- •Instant vision detection and context window estimation in heuristic stage
- •Custom classifier LLM as primary task tagger (task family, complexity, capabilities)
- •Heuristic fallback if classifier is unavailable
Fixed
IPv6/Happy Eyeballs Fix
- •Fixed Node.js 23 + Cloudflare ETIMEDOUT errors causing timeouts
- •dns.setDefaultResultOrder("ipv4first") applied at server startup
- •Affected providers: Groq, OpenRouter, Cerebras, Mistral