Checked against the OpenRouter public models feed. Prices are copied, not estimated.

The Best Model ForWhich AI model to use, from ten labs.

The best AI model for cheap high-volume automation

The same small task, thousands of times, where a miss is cheap to catch.

Short answer

Start with Claude Haiku 5.5 ($0.10 in · $0.50 out per million tokens). Even the quality setting stays in the cheap band. Haiku 5.5 is the strongest reviewed high-volume model in this guide: summaries, routing, and subagent steps. A flagship is the wrong purchase for bulk triage. A cheaper stop is GPT-6 Luna ($0.10 in · $0.50 out).

Quality

Claude Haiku 5.5

Anthropic · $0.10 in · $0.50 out

Even the quality setting stays in the cheap band. Haiku 5.5 is the strongest reviewed high-volume model in this guide: summaries, routing, and subagent steps. A flagship is the wrong purchase for bulk triage.

Claude Haiku 5.5 is listed at $0.10 input and $0.50 output per million tokens, with a 1,000,000-token context window.

Balanced

GPT-6 Luna

OpenAI · $0.10 in · $0.50 out

Luna is described as the cost-efficient model for classification and light agent steps. It accepts images and files, which matters if the queue is not pure text.

GPT-6 Luna is listed at $0.10 input and $0.50 output per million tokens, with a 1,050,000-token context window.

Cheapest sensible

DeepSeek V4 Flash 0423

DeepSeek · $0.09 in · $0.18 out

April V4 Flash is the cheapest reviewed text model that still has a huge window and open weights. It does not see images.

DeepSeek V4 Flash 0423 is listed at $0.09 input and $0.18 output per million tokens, with a 1,048,576-token context window.

Also on the feed

These match the factual limits of the task (modality, tools, context) and are not the reviewed picks. A row marked new has no guide note.

Questions

Why not the flagship if quality is selected?

Because the task is bulk. Quality here means the best model that is still sensible to call thousands of times. Astra and Fable fail that test on price.

What should I sample?

A hundred real inputs, including the ugly ones. Count the misses. The listed token rate does not include the cost of a bad label.