Show HN: Price Per Token – LLM API Pricing Data
A new site that visualizes price-per-token for major LLM APIs is praised for its clean, focused UI but criticized for limited coverage, occasional pricing errors, and ignoring complexities like caching, tiered pricing, and differing tokenization schemes. Commenters point to more comprehensive alternatives (such as OpenRouter, models.dev, and various JSON sources) and argue that meaningful cost comparison should factor in model quality, provider differences, and real-world task costs rather than raw token prices alone. Several people suggest extending such tools with benchmarking, historical pricing, and personalized calculators to help developers choose models and monitor changing costs.
Existing tools & discoverability
- Several commenters point out prior LLM price comparison tools (OpenRouter, llm-prices.com, Helicone, models.dev, llmprices.dev, etc.) and are surprised the author didn’t find them.
- Some say they now just use OpenRouter or similar services to check prices instead of vendor pages.
Scope, completeness, and “low effort” debate
- Strong criticism: site initially covers ~26 models from 3 big providers, omitting many popular ones (Mistral, Llama, Gemma, DeepSeek, Qwen, Groq, etc.) and prompt-cache pricing, leading some to call it “low effort” or “a mockup.”
- Others strongly defend the project: they value the simplicity, clear UI/graph, and see it as a useful starting point that can be iterated on.
- The author says they intentionally started small to gauge interest and plans to add many more models and cache pricing.
Token pricing complexity
- Multiple comments argue that “price per token” alone is misleading:
- Tokenizers differ between models; images and structured output can be billed differently.
- Providers have batch pricing, off-peak pricing, context-window-based pricing, “thinking” vs non-thinking token prices, tiering, and implicit/explicit caching.
- Same model via different providers can have very different prices; open models often vary widely in cost across hosts.
- Some suggest the right unit is “cost of a standardized task run” rather than per-token price.
Requested features & enhancements
- Cost calculator for custom input/output token counts and blended input/output metrics.
- Benchmarks or leaderboards joined with pricing to show “bang for buck,” possibly per endpoint / API shape.
- Periodic standardized tasks (summarization, coding) to estimate real query cost, with timestamps and historical trend tracking.
- Additional metadata: context length, modalities, cache pricing, provider, tiered pricing, etc.
- Monitoring/alerting on pricing changes as a potential paid service.
Data accuracy & maintenance
- One pricing error (Gemini 2.5 Flash Lite) is called out; the initial defensive response and later correction spur discussion about tone and trust.
- Several people discuss scraping APIs (e.g., OpenRouter, LiteLLM) and using agents/scrapers to keep a prices database continuously up to date.