Open-model per-host spread

The Open-model per-host spread is 1.9992 × (max/min) on 2026-10-23 (n=78 model/quantization groups). Definition: For open-weight models served by at least three hosts (OpenRouter endpoints and first-party host price pages), the median ratio between the highest and lowest listed input price across hosts serving the same quantization (1.0 = hosts agree).

Computed from the SKU state at the end of each UTC day (latest value: 2026-10-23; today's value is recomputed until the day ends). The constituents of the latest value are listed below; each shows when it was verified and links to its source.

Constituents of the latest value

Details of the latest value

open weight models

4141

unmatched models

225

min hosts

3

min coverage

3

first party skus

35

openrouter rows replaced by first party

21

community rows excluded

175

groups with min hosts

78

output price median spread

2.0383

{
 "first_party_hosts": [
  "BaseTen",
  "Cerebras",
  "Groq",
  "SambaNova",
  "Together"
 ],
 "widest": [
  {
   "hosts": 8,
   "max_usd_per_1m": 2.1,
   "min_usd_per_1m": 0.08,
   "model": "z-ai/glm-5.3",
   "quantization": "fp4",
   "ratio": 26.25
  },
  {
   "hosts": 8,
   "max_usd_per_1m": 0.99,
   "min_usd_per_1m": 0.04,
   "model": "qwen/qwen3.8-27b",
   "quantization": "unknown",
   "ratio": 24.75
  },
  {
   "hosts": 13,
   "max_usd_per_1m": 1.68,
   "min_usd_per_1m": 0.14,
   "model": "z-ai/glm-5.3",
   "quantization": "unknown",
   "ratio": 12.0
  },
  {
   "hosts": 10,
   "max_usd_per_1m": 0.45,
   "min_usd_per_1m": 0.039,
   "model": "qwen/qwen3.8-27b",
   "quantization": "fp8",
   "ratio": 11.5385
  },
  {
   "hosts": 6,
   "max_usd_per_1m": 3.0,
   "min_usd_per_1m": 0.26829,
   "model": "deepseek/deepseek-v3.2",
   "quantization": "unknown",
   "ratio": 11.1819
  },
  {
   "hosts": 4,
   "max_usd_per_1m": 0.44,
   "min_usd_per_1m": 0.0422,
   "model": "deepseek/deepseek-v4-flash-0731",
   "quantization": "fp4",
   "ratio": 10.4265
  },
  {
   "hosts": 12,
   "max_usd_per_1m": 0.225,
   "min_usd_per_1m": 0.022,
   "model": "z-ai/glm-5.3-flash",
   "quantization": "unknown",
   "ratio": 10.2273
  },
  {
   "hosts": 5,
   "max_usd_per_1m": 3.0,
   "min_usd_per_1m": 0.3,
   "model": "moonshotai/kimi-k3",
   "quantization": "fp4",
   "ratio": 10.0
  },
  {
   "hosts": 8,
   "max_usd_per_1m": 1.32,
   "min_usd_per_1m": 0.132,
   "model": "deepseek/deepseek-v4-pro-0813",
   "quantization": "fp8",
   "ratio": 10.0
  },
  {
   "hosts": 9,
   "max_usd_per_1m": 0.4092,
   "min_usd_per_1m": 0.04444,
   "model": "deepseek/deepseek-v4-flash-0731",
   "quantization": "fp8",
   "ratio": 9.2079
  }
 ]
}

Definition

For open-weight models served by at least three hosts (OpenRouter endpoints and first-party host price pages), the median ratio between the highest and lowest listed input price across hosts serving the same quantization (1.0 = hosts agree).

Inclusion rule

inference/openrouter/<author>/<slug>/<provider>/<quant> rows; model must be open-weight per llm/modelsdev (joined on the OpenRouter model id, then on the bare model slug; OPEN_MODEL_ALIASES for manual joins; unmatched models are counted); group by (model, quantization); endpoints in a healthy status and default service tier; price 0 (free) dropped; >= 3 distinct hosts per group; spread = max/min input price; headline = median over groups. The output-price variant, the ten widest spreads and the group count are published. Method change (wave 3): first-party host SKUs (groq, together, sambanova, cerebras, baseten: the host's own price page or API) join alongside the OpenRouter rows. They publish no quantization, so they are 'unknown' and are grouped by model only within the (model, unknown) group, never mixed into a fp8/bf16 group. A host that has a first-party SKU for a model counts once for that model, with its first-party price (its OpenRouter endpoints for that model are dropped). Community sources (the Hugging Face router, hf-router) never enter. A stale first-party source drops out; a stale OpenRouter or models.dev source voids the series.

When it cannot be computed: Not computed: that day shows no number and states the reason; the last full value stays shown with its date. A source whose last confirmed snapshot is older than 3x its cadence (6h sources 18 h, 12h sources 36 h, 1d sources 72 h) drops out; below minimum coverage the series is not computed and says 'not verified: <source> stale since <date>'.

Minimum coverage: 3 model/quantization groups. Below that the series is not computed and says so instead of showing a number.

Sources feeding it

openrouter-endpoints, modelsdev, groq-models, together-pricing, sambanova-models, cerebras-models, baseten-pricing

Deliberately excluded

• free variants (price 0)
• unhealthy or non-default-tier endpoints
• models that are not open-weight or cannot be joined
• hf-inference-providers (community cross-check)

What can break it

• missing or inconsistent quantization labels
• promotions
• a model changing alias
• hosts that list but do not serve
• OpenRouter changing its endpoint schema

Daily values, last 90 days

Every day since 2023-09-07: CSV · JSON.

API: JSON · CSV

Rack Rates (2026). Open-model per-host spread [data as of 2026-10-23]. https://rackrates.dev/index/open-model-host-spread (accessed 2026-10-23). List prices, not transacted prices. CC BY 4.0.

@misc{rackrates_index_open_model_host_spread_2026,
  author       = {{Rack Rates}},
  title        = {Open-model per-host spread},
  year         = {2026},
  howpublished = {\url{https://rackrates.dev/index/open-model-host-spread}},
  note         = {Data as of 2026-10-23; accessed 2026-10-23; list prices, not transacted prices; CC BY 4.0}
}