Qwen2.5 72B Instruct

qwen/qwen2.5-72b-instruct

Qwen (Alibaba)· 33K ctx

API onlytool callstructured quant: unknown

spread (output, p90 ÷ p10)

— too few providers

$0.590 – $0.590 · 1 providers · max÷min 1.0×

1 provider priced · 2 offerings

Including routers & resellers

Cheapest / median / spread use providers that serve the model on their own capacity: first-party lab, hyperscaler, inference host.

Provider classes in the stats:

Blended $/M assumptionsInput : output: 1Cache-hit share of input%

blended = 90.9% input × [(1 − 70% cache) × input + 70% cache × cache-read] + 9.1% × output + cache writes (30% of uncached input) where a write price is published

A provider that publishes no cache price is blended with cache share 0 (all input at the input price) and labelled “no cache price published” — never penalised against identical in/out prices.

Provider	Input	Output	Cache rd	Blended
Abacus router / aggregator	$0.110	$0.380	—	$0.135no cache price published
Kilo Gateway router / aggregator	$0.360	$0.400	—	$0.364no cache price published
NanoGPT router / aggregator	$0.360	$0.400	$0.178	$0.248
OpenRouter router / aggregator	$0.360	$0.400	—	$0.364no cache price published
SiliconFlow inference host	$0.590	$0.590	—	$0.590no cache price published

Connect

Provider (default: cheapest output among direct providers)

$0.590 /M · inference host

opencode.json

{
  "provider": {
    "siliconflow": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://api.siliconflow.com/v1"
      },
      "models": {
        "Qwen/Qwen2.5-72B-Instruct": {}
      }
    }
  }
}

Set SILICONFLOW_API_KEY in your environment. Generated — verify against the OpenCode docs.

Not enough history yet — the min/median/max sparkline fills in as the daily snapshot accumulates.

Data: models.dev (MIT), snapshot 2026-10-25. Prices are vendor-published; verify before purchase. Methodology.