The real prices of frontier models. Tokens * Price, right? (playcode.io)

Recent analysis reveals a hidden pricing discrepancy in AI model costs, particularly when builders utilize coding agents. Though pricing pages typically showcase a straightforward cost-per-token, the actual number of tokens generated from a given piece of text varies significantly across different models due to their unique tokenizers. For instance, a TypeScript file might consist of 681 tokens on GPT's tokenizer but balloon to 1,178 tokens on Anthropic’s newest model. This disparity exacerbates costs, as a model seemingly priced at $5.00 per million tokens could end up being much pricier depending on how many tokens your content converts into. This nuance is crucial for the AI/ML community, as it calls for more transparent comparisons between models. The analysis identified two major cost layers: first, a 'stealth hike' where updates to a model's tokenizer result in increased token counts without a corresponding price reduction; and second, the challenge posed to developers, particularly those working with code, as tokenization distortion significantly impacts their billing. The real implications are that builders must be vigilant when estimating costs, as conservative projections can lead to unexpected expenses—uplifting the necessity for comprehensive token metrics on all pricing pages to avoid costly surprises.