大模型网摘After Third-Party Hosting Price Hikes: A Cost Checklist for Moving Back to the Official API
On October 20, the AI industry newsletter stork.ai reported that Alibaba had "quietly" raised the output price of hosted DeepSeek V4 Flash 0731 by 230% — roughly 3.3 times the previous rate. As of press time there is no public explanation from the platform, so treat actual invoices as the source of truth. For teams routing most traffic through third-party hosting, this is the moment to redo the math: the convenience premium just got a lot more expensive.
Official API reference prices (October 2026)
The official DeepSeek API currently serves deepseek-flash (the old deepseek-v4-flash and v4-flash-vision-exp names are retired, with requests auto-served at Flash pricing): cache-hit input $0.003–0.006 per million tokens, cache-miss input $0.15–0.30, output $0.60–1.20 (halved during off-peak windows); 1M context and a concurrency cap of 2,500. Background in the V4.1-Flash-1017 release notes.
A six-step checklist
1. Audit model IDs and volumes: list every model served via third-party hosting, plus monthly input/output tokens and hit rates. 2. Build a two-way cost table: compare on the same basis, splitting cache-hit and cache-miss; assume a conservative hit rate after migration. 3. Switch the endpoint: the official API is OpenAI-compatible, so most cases only need a base_url and api_key change. 4. Regression-test: long-context behavior, vision input, and the default-on thinking mode deserve extra attention. 5. Turn on the savings: enable prefix caching for stable prompts and move deferrable batch jobs into off-peak hours — details in the peak/off-peak pricing explainer. 6. Keep a fallback: retain the old route in your gateway for one full billing cycle before decommissioning it.
Two traps to avoid
First, cache hit rate drives your effective unit price — the gap between hit and miss is roughly 50x, so prompt-template stability matters more than which route you use; see the cost-saving guide for the full playbook. Second, hikes like 230% tend to arrive without advance notice: the heavier your hosting dependency, the bigger your passive cost risk. Adding a monthly invoice review to the ops checklist is a cheap hedge.