70k DeepSeek requests a week: what I learned trying to forecast my LLM API bill

TL;DR: My cheap DeepSeek host’s promo prices look like they’re ending, so I priced my 70k requests a week elsewhere. Cheap per token isn’t cheap per task: one of the cheapest DeepSeek hosts costs more than twice as much per task as DeepSeek’s own API, prices change with days’ notice, and only 16% of my prompt tokens actually came from the cache. Budget at the price that applies after the promo, and measure cost per request on your own traffic. I run a pipeline that pushes text through an LLM around the clock, and how much arrives isn’t up to me. Last week that was about 70,000 requests, 124M prompt tokens and 12M output tokens, almost all on DeepSeek V4 Flash. This week one of the hosts I’ve been using told its users that model availability and pricing will change soon. I’ve been on its promo prices, so I’m guessing those are ending, and I’d rather plan for it now than find out from the invoice. At DeepSeek’s own list prices, last week would have cost about $28, or $46 if it had all run at peak hours. So I sat down to work out what else could move that number. Prices below are as of Oct 22. The first problem is that prices don’t sit still. DeepSeek gave three days’ notice before its August price increase, and its September switch to V4.1 Flash took effect the day it was announced. Google released three Flash models in six weeks, two of them at half price until Dec 31, so a newer one may replace 3.8 Flash before anyone pays its full price. Hosts move faster still. The cheapest GLM-5.3 Flash host on OpenRouter cut its input price by about 70% and more than doubled its output price within a day, then changed both again a few hours later. The price per token also doesn’t tell you how many tokens a model will use. Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 / $0.50, but on Artificial Analysis’s test run, both at Max, Haiku used three times the output tokens and cost $330 against $122. On Oct 17 DeepSeek swapped the V4.1 Flash build behind the same name: same price, 6 points higher score, 26% cheaper to run. Caching is where it gets sneaky. On paper it’s the biggest saving: when three quarters of a request is fixed instructions, the cache halves the bill on most models. In practice, only about 16% of my prompt tokens came from the cache last week. And Anthropic’s cache expires 5 minutes after last use and costs 1.25x input to write, so if requests come more than 5 minutes apart, 1,000 of them cost $0.63 against $0.55 with no cache at all. Gemini 3.8 Flash doesn’t cache prompts under 4,096 tokens. Then there’s the host. DeepSeek V4.1 Flash has 38 endpoints on OpenRouter, with output prices from $0.18 to $2.40 per 1M tokens. DeepInfra is one of the cheapest per token, but it ran at 68 tok/s against 278 on Together, and Artificial Analysis puts its cost per task at $0.61 against $0.26 on DeepSeek’s own API. In the worked example (10,000 requests a day, 4,000-token prompts, warm cache, thinking off), that comes to about $84 a month on Haiku 5.5 and $123 on DeepSeek spread over the week, or $203 if it all ran at peak rates. Gemini 3.8 Flash at full price comes to about $2,475 before any thinking tokens, because it can’t use the cache at this size. How do you forecast your API bill when prices and models keep shifting under you? And what has actually cut your costs: caching, batch, off-peak hours, switching hosts, a smaller model, or something else? Full post, with the worked example across 13 model/host combos and charts: easiapiai.com/articles/ai-costs-hard-to-project.html