Cloud Auto-Model Selection & Tier Aliases

Model Tiers & Pareto-Optimal Routing

To simplify model selection across hundreds of available LLMs, CBorg provides four curated tier aliases representing distinct points along the cost-vs-intelligence Pareto-optimal frontier (based on published industry benchmarks from Artificial Analysis):

Tier AliasCurrent Target ModelFrontier Position & ProfilePrimary Use Case
cloud-auto-budgetgpt-5.6-luna-medium-compressedMaximum Cost-Efficiency: Lowest token cost with rapid response latency.Basic Q&A, lightweight text formatting, data transformation, high-volume classification.
cloud-auto-standardgpt-5.6-luna-max-compressedBalanced Everyday Utility: Solid intelligence with moderate cost.General multi-turn chat, standard code generation, document drafting, summarization.
cloud-auto-advancedgemini-flash-high-compressedHigh Performance & Technical Depth: Strong reasoning at competitive speed.Complex scriptwriting, multi-step code synthesis, data analysis, math problem-solving.
cloud-auto-maxgpt-5.6-sol-max-compressedFrontier Intelligence: Deep reasoning and maximum analytical capability.Architecture design, difficult debugging, formal mathematical proofs, complex logic puzzles.

Info

Dynamic Frontier Updates: As new models are released, benchmarks evolve on Artificial Analysis, and pricing structures change across providers, we will periodically update the underlying model mappings for these aliases. By pointing your applications to cloud-auto-* aliases, your workloads automatically benefit from ongoing frontier improvements without requiring code changes.

All tier aliases are automatically paired with Headroom Context Compression (headroom-compressed) to compress conversational and tool-call history, maximizing usable context and reducing prompt latency.

Automated Selection: cloud-auto-adaptive

If you prefer not to manage tier selection manually per request, CBorg offers the cloud-auto-adaptive smart router.

cloud-auto-adaptive dynamically analyzes each incoming prompt in real-time using LiteLLM’s complexity classification engine. It evaluates prompt characteristics—including input token count, structural complexity, reasoning cues, and coding keywords—and automatically dispatches the query to the optimal Pareto tier:

  • Simple prompts route to cloud-auto-budget.
  • Standard conversational queries route to cloud-auto-standard (the default fallback).
  • Technical analysis and coding tasks route to cloud-auto-advanced.
  • Deep reasoning challenges route to cloud-auto-max.

Usage Examples

Using cURL with cloud-auto-adaptive

curl https://api.cborg.lbl.gov/v1/chat/completions \
  -H "Authorization: Bearer $CBORG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cloud-auto-adaptive",
    "messages": [
      {
        "role": "user",
        "content": "Write a Python script that parses server access logs, computes p95 response times, and outputs a formatted markdown report."
      }
    ]
  }'

Using Python (OpenAI SDK) with a Specific Tier Alias

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ.get("CBORG_API_KEY"),
    base_url="https://api.cborg.lbl.gov/v1",
)

# You can request the smart router 'cloud-auto-adaptive' or a specific tier alias
response = client.chat.completions.create(
    model="cloud-auto-advanced",
    messages=[
        {"role": "system", "content": "You are a scientific computing assistant."},
        {"role": "user", "content": "Explain how Monte Carlo simulations can estimate pi, with sample code."},
    ],
)

print(response.choices[0].message.content)