Artificial intelligence news.
Straight from the original announcements.

OpenRouter calculates when self-hosting an LLM beats API costs

November 12, 2026By the AI News Bites Editorial TeamSource: OpenRouter
OpenRouter calculates when self-hosting an LLM beats API costs
Image: OpenRouter

Self-hosting an open-weight LLM costs less than using an API only when traffic keeps the GPU busy enough, according to an OpenRouter analysis. Its calculations show that a Qwen3.8 27B deployment can break even at 6 to 12% load on one rented RTX 3090, while some DeepSeek V4.1 Flash setups never beat the cheapest zero-data-retention API endpoint.

Key points
  • Self-hosting becomes cheaper only when GPU utilization passes the break-even load.
  • Qwen3.8 27B on one rented RTX 3090 breaks even at 6 to 12% load in OpenRouter's examples.
  • DeepSeek V4.1 Flash on four H200s never beats the cheapest ZDR API endpoint.
  • An owned RTX 3090 reaches a 1% break-even point when only power costs are counted.
  • OpenRouter says ZDR endpoints and regional domains cover many privacy requirements without self-hosting.

The break-even calculation

OpenRouter compares one hour of GPU time in two ways: the hourly rental cost and the API price for the tokens that the same hardware can serve. Its formula is break-even load = GPU rent per hour / API cost of one fully loaded GPU-hour of tokens.

A break-even point of 30% means the GPU must be busy for 30% of the rented hours before self-hosting costs less than the API. A result above 100% means the API remains cheaper even if the GPU never sits idle. OpenRouter says its calculations exclude engineering time and spare capacity for traffic peaks, making every break-even figure a lower bound.

The comparison also needs the same token mix on both sides. Input, cached input and output tokens have different API prices, and reasoning models can generate substantial output before producing an answer.

Hardware, models and API prices vary widely

OpenRouter used API prices from October 8, 2026, and compared multiple providers serving the same open-weight models. Its model endpoint data listed 18 endpoints for Qwen3.8 27B, including 14 with zero-data-retention policies, and 31 for DeepSeek V4.1 Flash, including 24 ZDR endpoints.

For Qwen3.8 27B, input prices ranged from $0.024 to $0.99 per million tokens, while output prices ranged from $1.49 to $5.25. DeepSeek V4.1 Flash input prices ranged from $0.016 to $0.45, and output prices from $0.14 to $1.80. OpenRouter says the most expensive provider charged 3.5 times the cheapest provider for Qwen3.8 27B output, and nearly 13 times the cheapest price for DeepSeek V4.1 Flash.

The self-hosted assumptions included DeepSeek V4.1 Flash on four H200 GPUs at $2.90 or $4.59 per GPU-hour, and on two B200 GPUs at $3.70, $6.79 or $6.89 per GPU-hour. Qwen3.8 27B was assessed on one RTX 3090 at $0.50 per hour, or with power costs of about $0.05 per hour for an owned card.

Qwen3.8 27B reaches break-even at low load

Against the cheapest ZDR endpoint, Qwen3.8 27B on a rented RTX 3090 reached an 8% break-even load with short prompts and 12% with OpenRouter's estimated chat mix. Against the median ZDR price, the figures were 6% and 8%. Using an owned RTX 3090 and counting power only reduced the break-even point to 1%.

OpenRouter estimates that a rented 3090 costs $365 per month and pays for itself after about 170 million output tokens of its chat mix. However, the setup served 64 concurrent requests at about 15 tokens per second per user. The analysis says that speed is slow for a chat interface, particularly for a reasoning model that produces thousands of hidden tokens before its answer.

The RTX 3090 result is also described as a best case because it uses HyperQwen's heavily patched vLLM build. OpenRouter used its end-to-end result, including prefill, rather than its 1,035 tokens-per-second decode-only figure.

Large deployments need sustained traffic

DeepSeek V4.1 Flash produced a different result. The four-H200 setup never beat the cheapest ZDR endpoint, while the two-B200 setup required 76% load against that endpoint at the lowest listed rent. Against the median ZDR price, the two-B200 setup reached 37% load at $3.70 per GPU-hour, but the higher-rent configurations did not always break even.

OpenRouter says traffic limited to working hours, 40 of the week's 168 hours, can keep a GPU busy at most 24% of the time. At the October 8 prices, that ruled out every DeepSeek setup and left the RTX 3090 option. A batch pipeline running around the clock could reach 70% load, bringing both B200 rows into range against the median ZDR price, but neither against the cheapest endpoint.

The analysis also says that most privacy requirements do not require self-hosting. OpenRouter's provider.zdr setting limits a request to zero-data-retention endpoints, while its EU and US domains keep requests in-region. Prices can change quickly: on October 19, a new ZDR endpoint for Qwen3.8 27B charged $0.50 per million output tokens, moving the short-prompt break-even to about 29%. Users therefore need to check current prices, expected traffic and acceptable response speed before buying or renting GPUs.

Related news

OpenRouter finds no clear quality gap between fp8, fp4 and fp16
ResearchNovember 11, 2026

OpenRouter finds no clear quality gap between fp8, fp4 and fp16

OpenRouter found no detected accuracy difference between fp8, fp4 and fp16 in a test of Qwen3.8 27B, although the company cautions that the result covers one model and…

vLLM Expands Encoder Disaggregation for Multimodal AI Serving
For developersNovember 11, 2026

vLLM Expands Encoder Disaggregation for Multimodal AI Serving

vLLM has detailed engineering work that separates multimodal encoding from language-model prefill and decode, allowing operators to scale those stages independently. The…

Open-weight models hit 72% of top-50 token use on OpenRouter
ResearchOctober 31, 2026

Open-weight models hit 72% of top-50 token use on OpenRouter

Open-weight models handled 72% of the tokens among OpenRouter's top 50 models in late September 2026, up from 27% in October 2025, according to a new analysis from the…