A new site called CostPerPrompt just launched, and it does one thing most pricing pages don’t: it turns raw token rates into real dollar answers. According to Hacker News, where the Show HN post climbed to 167 points, the tool tracks live pricing for 232-plus AI models, refreshes automatically, and layers on calculators built around actual workloads instead of abstract per-token math.
What stands out here is the framing. Most “how much will this cost” articles quote a headline price and stop. CostPerPrompt models the messy reality: growing conversation history, retries, cache hits, and batch discounts. That’s the difference between an estimate that’s roughly right and one that’s off by 2 to 3 times.
What it actually does
Hacker News describes a spread of calculators, each aimed at a specific real-world workload:
- API Cost Calculator. Compare any workload across all 232 models, with prompt caching and batch discounts factored in. This is the core engine the rest builds on.
- Chatbot Cost. Simulates real conversations, including history that grows over time and context that gets resent every turn. That resent context is where chatbot bills quietly balloon.
- Agent Cost. Multi-step loops, tool schemas, and retries. The site’s blunt claim: agents cost 10 to 30 times more than people expect. If you’re building agentic workflows, that number alone is worth a look.
- RAG Cost. Prices indexing, retrieval, and generation separately. The takeaway, per the site, is that embeddings are pennies, so the real cost sits elsewhere in the pipeline.
- Voice AI Cost. Breaks the bill into its three parts, speech-to-text, the LLM, and text-to-speech, priced per minute, per call, and per month.
- GPU Rental Pricing. H100, A100, and RTX 4090 rates across 10 providers. The eye-opener: the same chip can carry a 5 times price spread depending on who you rent it from.
- Image API Pricing. From $0.003 to $0.17 per image, sorted into three market tiers, plus a batch calculator.
- Token Counter. Paste text, get the token count and cost on any model. It runs entirely in your browser, so nothing you paste leaves your machine.
The pricing logic underneath
The site includes a plain-English primer on how AI billing works, and it’s a useful refresher. Every major provider charges per token, roughly three-quarters of a word, with separate rates for input (what you send) and output (what the model writes back). Output usually runs 3 to 5 times more expensive than input. So a model listed at $5 / $25 per million tokens charges $5 for every million you send and $25 for every million it generates.
Two discounts change the math dramatically, and this is where the calculators earn their keep. Prompt caching cuts repeated input costs by up to 90 percent, which matters enormously for chatbots that resend conversation history on every turn. Batch processing takes about 50 percent off when you can afford to wait for results. The site says most cost estimates ignore both, which is exactly why they run so far off.
Why it matters
This is significant because cost is becoming the deciding factor in model selection, not just capability. When a chatbot, an agent, and a RAG pipeline each bill differently, and when the same H100 swings 5 times in price between providers, guesswork gets expensive fast. A tool that models your specific workload before you commit is the kind of thing finance teams and solo builders both need.
The honest caveat: these are calculators, not invoices. Real usage patterns vary, and live pricing pages are only as accurate as their last refresh. Treat the outputs as strong directional estimates, not guarantees. The site doesn’t mention pricing for itself, and the browser-based token counter suggests the free-to-use tools are the draw.
For anyone budgeting an AI feature right now, plugging in your real numbers before you build beats discovering the bill after you ship. You can find the full set of calculators at the original source.