LLMs vary wildly in quality by task. GPT-4o dominates coding, Gemini excels at long context, DeepSeek leads math reasoning. ArcRouter picks the right model for every prompt — automatically.

• Hybrid semantic routing — lexical prefilter + benchmark shortlist + embedding reranking
• 24 topic categories with subcategory detection (code/frontend, math/calculus, science/physics, ...)
• 4 complexity tiers — SIMPLE, MEDIUM, COMPLEX, REASONING
• Agentic detection — auto-detects tool-use prompts, routes to tool-capable models
• 345+ models scored with real benchmarks (LiveBench, LiveCodeBench, HumanEval, GPQA)

• Council mode — query 3-7 models in parallel, return the consensus answer
• Semantic grouping — embedding similarity (paid) or Jaccard overlap (free) to find agreement
• Chairman escalation — when models disagree, a synthesis model resolves conflicts

• X-Agent-Step header — hint complexity per step (simple-action, code-generation, reasoning, verification)
• Workflow budgets — set a total USD budget per workflow, auto-downgrade at 60/80/95% spend
• Session model pinning — lock a model for a conversation via session_id
• Workflow telemetry — GET /v1/workflow/{session_id}/usage for spend, latency, tier distribution

• Prompt compression — lossless L1 (dedup) + L2 (whitespace) + L5 (JSON compaction), 10-15% token savings
• 5 direct providers — OpenAI, Anthropic, Google, DeepSeek, xAI (no middleman markup)
• OpenRouter fallback — 300+ long-tail models when direct access isn't available
• Max cost filtering — max_cost per request, graceful downgrade to cheapest viable model
• Routing metadata — every response includes estimated_cost_usd, savings_vs_gpt4_pct, models_considered

• Circuit breaker + fallback chains — auto-retry with up to 3 fallback models on failure
• Context-length filtering — excludes models with insufficient context window
• Model exclusion — exclude_models to block specific models per request

• x402 micropayments — pay per request with USDC on Base, no API key needed
• Stripe metered billing — traditional API key billing for teams
• Free tier — no auth required, free models, rate-limited

• Model aliases — "claude", "gpt", "gemini", "deepseek" instead of full model IDs
• OpenAI-compatible API — drop-in replacement, same /v1/chat/completions format
• Streaming — SSE streaming with routing metadata in the first chunk
• Usage tracking — GET /v1/usage for per-key daily stats

# TypeScript SDK
pnpm add @arcrouter/sdk

# MCP server for Claude Code / Cursor / Cline
claude mcp add arcrouter --transport http https://api.arcrouter.com/mcp

# Direct API (no auth required for free tier)
curl https://api.arcrouter.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"Hello"}],"budget":"auto"}'

import { ArcRouter } from '@arcrouter/sdk';

const arc = new ArcRouter();
const res = await arc.chat('Explain quantum entanglement');

console.log(res.content);           // The answer
console.log(res.routing.model);     // e.g. "gemini-2.0-flash"
console.log(res.routing.topic);     // "science/physics"
console.log(res.routing.complexity);// "MEDIUM"

Prompt → Classify (topic + complexity + agentic) → Score 345+ models → Semantic rerank → Route → Fallback chain

1. Classify — detect topic (24 categories), complexity (4 tiers), agentic intent, and confidence
2. Score — rank models using real benchmarks, weighted by semantic relevance + value + reliability
3. Route — hybrid semantic routing narrows to optimal model, filtered by budget + context length + cost
4. Respond — direct provider call with circuit breaker, up to 3 fallback models on failure
5. Compress — lossless prompt compression on long conversations (>5000 chars) to reduce token cost

• Website: arcrouter.com
• API: api.arcrouter.com
• Docs: arcrouter.com/docs
• npm: @arcrouter/sdk · arcrouter-classifier