A month ago I published a post about burning €11 in 18 minutes on GreenPT. The session logged 10 million input tokens with zero cache hits. Every request billed the full conversation history.
GreenPT support responded to that post and refunded the €11. They said they do caching internally, but that it’s opaque from the API — you wouldn’t be able to see cache stats in the API response.
That’s no longer the case. Their pricing page now lists cached input rates for several models, and thecached_tokensfield shows up in the API response. I tested both models directly against the GreenPT API last night.
deepseek-v4-flash-0731with a ~270-token system prompt, three calls back to back:
Call 1: 274 prompt, 0 cachedCall 2: 274 prompt, 256 cached (93%)Call 3: 274 prompt, 256 cached
glm-5.2 — the model that burned me the first time — same test:
Call 1: 288 prompt, 0 cachedCall 2: 288 prompt, 256 cached (89%)Call 3: 288 prompt, 256 cached
In a real OpenCode session with DeepSeek Flash (17 requests), the cache hit rate was 87.9%. The session cost €0.027. Without caching it would’ve been €0.07.
I don’t know what changed. My OpenCode config hasn’t. GreenPT support didn’t mention any fixes on their side. But it’s working now.
Next: I’m going to test DeepSeek V4 Flash on GreenPT for my regular coding work, instead of running models locally on my MacBook. The goal is more computing power than a fully local model can give me, while keeping environmental impact as low as I can. At €0.04/M for cached input with a ~90% hit rate, the cost works too.