Why a 27B model costs more per token than much bigger ones

The 27B model is dense, and the "bigger" models it loses to are mixture-of-experts (MoE) models. On these models you pay for the parameters that run on each token, not for the total size. Here are today's OpenRouter prices, in USD per million tokens:

Qwen's naming shows the split. "27b" with no -aXb suffix means all 27B weights run on every token. "35b-a3b" means 35B weights in total, but only about 3B are active per token. Even inside the Qwen family, the 35B MoE costs half as much as the 27B dense model for output. The earlier dense 27B models were priced the same way: qwen3.5-27b at 2.60 output and qwen3.6-27b at 3.20.

1. Output tokens are where dense models cost more.

• Input is cheap at 0.15, but output costs 12.5 times as much.
• When generating, the provider has to read every active weight from GPU memory for each new token.
• A dense 27B model reads 27B weights per token. An A3B MoE model reads about 3B, so it can serve far more tokens per GPU-second.
• Processing the input happens in large parallel batches, which is why the input price stays low.

2. Fewer providers, so less price competition.

• qwen3.8-27b has 21 providers on OpenRouter. deepseek-v4.1-flash has 40, and that many providers competing has pushed its price down.
• The 1.875 output price shown on the model page is DeepInfra's bf16 endpoint, the full-precision default. It is not the cheapest one.
• The cheaper qwen3.8-27b endpoints are all quantized (lower-precision weights), which can reduce output quality:
  • Chutes (fp8): 0.039 / 1.25
  • Near AI (fp8): 0.04 / 1.35
  • Darkbloom (fp4): 0.035 / 1.50
• If you need it cheaper, pin a provider with OpenRouter's provider.order setting, or sort by price. Check that the quantization is acceptable for your use first.

3. A long context window adds cost. The 1M-token context endpoints (Novita, Alibaba, Phala) cost more than the 262K ones. A dense model's memory for past tokens per sequence is also typically larger than that of an MoE model with few active parameters. That means fewer requests fit in one batch, and each token costs more to serve.

Bottom line: a dense 27B model only makes sense on cost if you can run it yourself on one GPU, which is the reason dense models of this size exist. Through a hosted API, an MoE with few active parameters will almost always be cheaper per token. That's why your Baseten agent profile uses DeepSeek-V4.1-Flash at 0.30 / 1.20, which already undercuts qwen3.8-27b's default price.