Cheapest frontier models per 1M tokens
Published 2026-10-17 · Updated 2026-10-26
Price-per-token comparisons lie in boring ways, so a few ground rules first: always compare input and output separately (output is usually 3–10× input), check whether cached-input pricing exists, and remember that a chatty model that writes more tokens to reach the same answer costs more than the table implies.
Reading the OpenRouter pricing fields
The /models endpoint prices in dollars per token. Multiply by a million for the number everyone quotes. prompt is input, completion is output, and input_cache_read, when present, is the discounted price for re-sent context.
The cheap end of the frontier
As of late October 2026, the pattern across the catalog:
- Chinese labs dominate the cheap tier. DeepSeek, GLM (Z.ai), Kimi and MiniMax flagship models sit well under a dollar per million input tokens, often a few dollars on output.
deepseek-v4-class models are routinely the cheapest way to get near-frontier quality. - Mini and flash tiers are an order of magnitude cheaper than flagships.
gpt-6.1-mini,gemini-3-flash,claude-haiku-class models are where volume workloads belong. - Flagship pricing is bimodal. OpenAI and Anthropic top models run into double-digit dollars per million output tokens; the Chinese flagships deliver a large fraction of the capability at a fraction of that.
- Cache-aware pricing matters more than sticker price. For agentic workloads that re-send long contexts, a model with cheap
input_cache_readbeats a nominally cheaper model without it.
A fair way to compare for your workload
- Estimate your real input:output ratio, chat apps are often 4:1 input-heavy, summarization 10:1, agents 20:1.
- Price a representative request, not a token.
input_tokens × prompt + output_tokens × completion. - Add the cheapest-provider tax: rate limits and queueing cost engineering time too.
The catalog on this site carries the per-token numbers for each model, check the model page for the exact figures before committing a workload to a provider.