Strategic guide
Model routing cascade: When to use a cheap model vs a frontier model
The price gap between a budget model and a premium frontier model can be 50× or more per token. Most requests do not need frontier intelligence. A cascade strategy routes each task to the cheapest adequate tier and escalates only when necessary, cutting costs by 40–60% while preserving output quality.
Tiers
Four pricing and capability tiers
Tiers are computed from the current catalog. Prices change; verify the latest rates on model pages.
Budget tier — under $0.30 / 1M tokens
Best for: classification, extraction, simple Q&A, high-volume throughput
Fast and inexpensive. These models handle straightforward tasks where a wrong answer is easy to detect and retry. Ideal for filtering, labeling, and single-turn chat that does not need deep reasoning.
| Model | Provider | Input / 1M | Output / 1M | Context |
|---|---|---|---|---|
| Llama-3.2-3B-Instruct deepinfra | deepinfra | $0.020 | $0.020 | 131.1K |
| llama3.2-11b-vision-instruct lambda_ai | lambda_ai | $0.015 | $0.025 | 131.1K |
| llama3.2-3b-instruct lambda_ai | lambda_ai | $0.015 | $0.025 | 131.1K |
Mid tier — $0.30 to $2.00 / 1M tokens
Best for: RAG, summarization, customer-facing chat, moderate reasoning
Balanced capability and cost. Mid-tier models handle multi-turn dialogue, retrieval-augmented generation, and document summarization where context windows of 8K–32K are common. The extra cost over budget buys better instruction following and lower error rates.
| Model | Provider | Input / 1M | Output / 1M | Context |
|---|---|---|---|---|
| amazon.nova-lite-v1:0 bedrock_converse | bedrock_converse | $0.060 | $0.240 | 300.0K |
| zephyr-7b-beta anyscale | anyscale | $0.150 | $0.150 | 16.4K |
| gemma-7b-it anyscale | anyscale | $0.150 | $0.150 | 8.2K |
Premium tier — $2.00 / 1M tokens and above
Best for: complex reasoning, coding, analysis, agentic orchestration
Frontier models with the largest context windows, strongest instruction following, and broadest capability sets (tool use, code generation, structured output). Use these when task quality directly affects the user experience or when a budget/mid model repeatedly produces inadequate results.
| Model | Provider | Input / 1M | Output / 1M | Context |
|---|---|---|---|---|
| amazon.titan-text-premier-v1:0 Bedrock | Bedrock | $0.500 | $1.50 | 32.0K |
| CodeLlama-34b-Instruct-hf anyscale | anyscale | $1.00 | $1.00 | 4.1K |
| CodeLlama-70b-Instruct-hf anyscale | anyscale | $1.00 | $1.00 | 4.1K |
Reasoning tier — chain-of-thought models
Best for: math, logic, multi-step planning, self-critique chains
Models flagged as reasoning-capable produce internal chain-of-thought before answering. They excel at problems that require verification, backtracking, or step-by-step decomposition. Prices vary widely because the reasoning flag spans both budget and premium models.
| Model | Provider | Input / 1M | Output / 1M | Context |
|---|---|---|---|---|
| qwen3-4b-fp8 novita | novita | $0.030 | $0.030 | 20.0K |
| google.gemma-4-e2b Bedrock Mantle | Bedrock Mantle | $0.040 | $0.080 | 128.0K |
| gpt-oss-20b OpenRouter | OpenRouter | $0.020 | $0.100 | 32.8K |
Cascade strategy
How a cascade reduces cost
1. Route by task type
Classify each request: is it simple Q&A, RAG, coding, or planning? Assign the cheapest adequate tier.
2. Try the cheapest tier first
Send the request to a budget or mid model. Measure confidence or check output quality.
3. Escalate on failure
If the cheap model produces a low-confidence, incomplete, or incorrect result, retry with the next tier.
4. Track and tune
Log escalation rates per task type. Adjust tier assignments as new models appear or workload patterns shift.
Cost comparison
Cascade vs always-premium
In this scenario (2,000 input + 1,000 output tokens, 100,000 requests/month), a cascade routes 90% of requests to budget models, 8% to mid-tier, and 2% to premium.
Estimated savings: 95% on this workload. Your actual savings depend on task mix, escalation rate, and model-specific pricing.
Task type recommendations
Which tier fits your workload
| Task type | Recommended tier | Why |
|---|---|---|
| Classification, extraction, simple Q&A | Budget | High throughput, low complexity |
| Customer-facing chat, RAG, summarization | Mid | Balanced capability and cost |
| Coding, analysis, multi-step reasoning | Premium or Reasoning | Needs wider context and stronger logic |
| Math, logic, self-critique chains | Reasoning | Chain-of-thought output dominates |
| Multi-step agent (uncertain complexity) | Cascade (Budget / Mid / Premium) | Start cheap, escalate on failure |
Implementation patterns
Three ways to implement a cascade
Simple switch
Use an if-else block by task type. Each task type maps to one fixed model tier. Easy to implement and debug, but the tier assignment never changes per request.
Confidence-based
Call the budget model first. If the output confidence score (or a validation check) falls below a threshold, retry with a higher tier. Adapts per request at the cost of latency on failures.
Parallel
Send the same request to two tiers simultaneously. Pick the best result or use the premium output if the budget result fails validation. Lowest latency but double the token cost on every request.
Put your actual workload into the calculator
Token counts, request volume, and cache behavior change the optimal tier. Use the calculator with your numbers before committing to a routing plan.
Caveats
What this guide does not cover
This page defines tiers by price and capability flags only. It does not rank output quality, latency, rate limits, or provider reliability within a tier. The cascade pattern assumes your application can detect and retry low-quality outputs, which not all workflows support. Prices and model availability change frequently. Always verify the latest rates on model detail pages and official provider sources before building a production routing plan.