How a Missing Config Line Cost Me 38x More for the Same Model

Dmitry Platonov — Fri, 22 May 2026 19:32:07 +0000

I'm building a custom coding agent harness that uses DeepSeek models through OpenRouter. The models offer a good price/performance balance (especially with the current Pro discount).

While editing the config, I accidentally removed the provider preference. My OpenRouter workspace had only two providers enabled: the Official and ProviderA. And, of course, OpenRouter picked ProviderA! My harness displays session cost in real time, and within about 30 minutes the session cost jumped up enough to catch my attention. I pulled the usage breakdown:

Model	Provider	Calls	Prompt tokens	Cached tokens	Cache %	Completion tokens	Cost
flash	ProviderA	115	3,150,998	2,790,656	88.6%	66,176	$0.1471
flash	Official	718	28,198,126	27,187,072	96.4%	378,485	$0.3236
pro	ProviderA	11	306,246	93,952	30.7%	8,796	$0.3994
pro	Official	1341	54,105,295	51,849,600	95.8%	737,680	$1.8110

Normalized per 1M total tokens:

Flash: ProviderA $0.046, Official $0.011 -> 4x more expensive
Pro: ProviderA $1.27, Official $0.033 -> 38x more expensive

Why a "small" cache gap hurts so much

Prompt caching makes uncached input tokens an order of magnitude pricier (two orders in case of Official provider). Even the 8-point cache difference for Flash (88.6% vs 96.4%) means ProviderA processed 3x more uncached tokens. Combine that with a higher base price and you get the 4x multiplier. For Pro the cache gap is extreme (30.7% vs 95.8%) -- that's the core of the 38x blow-up.

Currently, official pricing for DeepSeek V4 Pro cached tokens is just $0.0036/M (yes, that's right, decimal point is in the right place!). For agentic workloads it massively drives cost down.

The fix: explicitly set provider preference

Always include the provider object in your HTTP request:

{
  "model": "deepseek/deepseek-v4-pro",
  "messages": [...],
  "provider": {
    "order": ["deepseek"],
    "allow_fallbacks": false
  }
}

Or at least create separate workspace/key guardrails on the OpenRouter for each model family.

Did something like that ever happen to you?

DEV Community: Dmitry Platonov

How a Missing Config Line Cost Me 38x More for the Same Model

Why a "small" cache gap hurts so much

The fix: explicitly set provider preference