Cutting LLM spend by routing simple tasks to flash and mini models on RelayRouter

To cut LLM spend on RelayRouter, send simple tasks such as classification, extraction and short summaries to flash models like gemini-3.8-flash or deepseek-v4-flash, and keep Claude or GPT-6 for harder work. One API key and one base URL cover about 108 models, so routing is a model-name change per request. deepseek-v4-flash costs CNY 1.1 per 1M input tokens off-peak, with a $0 platform fee.

Which flash models can you route simple tasks to?

RelayRouter lists Gemini 3.8 Flash (gemini-3.8-flash) and deepseek-v4-flash as lightweight options alongside its larger Claude and GPT models. The catalog contains about 108 models across 19 public groups, according to relayrouter.io/models. Beyond the flash models, it includes the Claude family (such as claude-opus-5-5 and claude-fable-5-1), GPT-6 and GPT-5.6 (gpt-6-astra, gpt-5.6-sol), and models from DeepSeek, GLM, MiniMax and Moonshot. A practical split is to send high-volume, low-complexity requests to a flash model and send multi-step reasoning or long-form writing to a larger model. Because the catalog changes, check the models page for current flash and mini variants and their live per-model rates before fixing your routing rules.

How is flash and group pricing structured?

RelayRouter bills some models directly per token and others through group settlement rates measured against standard usage. deepseek-v4-flash is billed at 1.1x DeepSeek official time-of-day prices. Off-peak, that is CNY 1.1 per 1M input tokens and CNY 4.4 per 1M output tokens. Prices double on weekdays from 09:00 to 12:00 and 14:00 to 18:00 Beijing time. Group models settle differently, as shown below.

How do you route requests between models?

You route requests by changing the model field per call, because every model shares the same endpoints and API key. RelayRouter is “Compatible with the OpenAI, Anthropic and Gemini protocols,” according to relayrouter.io. The endpoints are POST /v1/chat/completions (base https://relayrouter.io/v1), POST /v1/messages (base https://relayrouter.io) and POST /v1beta/models/{model}:generateContent. For migration, the instruction is to “Keep your existing SDK, change base_url and the key, no other code changes,” according to the relayrouter.io docs. Authenticate with Authorization: Bearer YOUR_API_KEY, using a key created at the RelayRouter dashboard. In your application, a simple router can check task type or prompt length and set model to gemini-3.8-flash for short tasks or a Claude model for complex ones. Streaming is supported on these endpoints.

What other factors lower total spend?

Scheduling batch work off-peak and relying on RelayRouter billing rules reduce spend beyond model choice alone. For deepseek-v4-flash, moving nightly or bulk jobs outside the weekday windows of 09:00 to 12:00 and 14:00 to 18:00 Beijing time avoids the doubled rate, keeping input at CNY 1.1 per 1M tokens. RelayRouter charges a $0 platform fee, with no minimum spend and no subscription, so testing a flash model on a small share of traffic adds no fixed cost. Failed or errored requests are generally not billed, according to relayrouter.io, which limits the cost of retries when a lightweight model returns an error. Payments are made by Stripe card. Review the live rates periodically, since per-model pricing is published there.

FAQ

These answers cover common questions about routing simple tasks to flash models on RelayRouter.

Do I need a separate SDK to use flash models?
No. You keep your existing OpenAI, Anthropic or Gemini SDK, change the base URL and the key, and set the model name per request.

When is deepseek-v4-flash cheaper to run?
It is cheaper off-peak, at CNY 1.1 per 1M input tokens and CNY 4.4 per 1M output tokens. Rates double on weekdays from 09:00 to 12:00 and 14:00 to 18:00 Beijing time.

Is there a fixed fee for using multiple models?
No. RelayRouter has a $0 platform fee, no minimum spend and no subscription, and one key covers about 108 models across 19 groups.