Llm API Cost Comparator Calculator

Comparing language model providers is a two-rate problem rather than a single-price one. Every request carries an input charge for the prompt and an output charge for the completion, and the two differ by a factor that is itself provider-specific: many models price output at four to five times input, while some hosted open-weight models such as Llama 3.1 405B price the two almost equally. That means the cheapest model for a workload depends on its shape. A summarization job with a 10,000-token document and a 200-token summary is dominated by the input rate; a drafting job with a short instruction and a long reply is dominated by the output rate, and the ranking of providers can invert between the two. The spread across a single comparison table is extreme - roughly two hundredfold between the least and most expensive entries on input tokens - which makes model routing the largest available saving in most systems. Sending classification, extraction, and routine question answering to a small model and reserving a frontier model for genuine reasoning typically cuts spend far more than prompt tuning does. Several caveats limit a pure price comparison. Rates say nothing about quality, latency, or rate limits, and a cheap model that needs two attempts is not cheap. Context window size does not change the per-token rate but does bound how much can be sent, and some providers apply a surcharge above a threshold such as 128,000 tokens. Batch endpoints, prompt caching, and volume agreements can move effective rates well below list.

Per-token metering became the industry norm with the OpenAI API in June 2020, which offered GPT-3 through four tiers named Ada, Babbage, Curie, and Davinci at prices quoted per thousand tokens. The structure was unusual for its time - most machine learning services billed per request or per compute-hour - but it matched the underlying economics, where cost scales with sequence length rather than call count. It also created the first genuinely comparable unit across vendors once competitors appeared. A competitive market formed rapidly in 2023. Anthropic made Claude commercially available that March, Google launched the PaLM API the same month and succeeded it with Gemini in December, and Mistral AI, founded in Paris in April 2023 by Arthur Mensch, Guillaume Lample, and Timothee Lacroix, opened its platform that December. In parallel, Meta released Llama 2 with open weights in July 2023, which let inference specialists host the same model at competing rates and introduced price competition on identical weights. Steep reductions followed as inference was optimized and hardware utilization improved, and around 2024 the industry converged on quoting dollars per million tokens instead of per thousand, largely because the leading rates had fallen far enough that the older unit produced unwieldy decimals. Cross-provider comparison tools exist because that convergence made the numbers directly comparable for the first time.