First token counts reveal Opus 4.7 costs significantly more than 4.6 despite Anthropic's flat pricing

Editor's take

Anthropic's latest Claude 3 Opus model, version 4.7, introduces a more granular tokenizer that increases the token count for identical input by up to 47% compared to Opus 4.6.

This change effectively raises the operational cost for users without altering the advertised per-token pricing. For businesses and developers heavily reliant on Anthropic's API, particularly those with high-volume usage or fixed budget constraints, this presents an immediate financial impact, potentially shifting cost-benefit analyses for model selection. It also highlights the opaque nature of tokenization and its direct influence on LLM economics.

Future observations should focus on whether other major LLM providers, such as OpenAI with its GPT-4 Turbo, will adopt similar tokenization strategies to manage computational costs, and what Anthropic's long-term approach will be to communicating these efficiency shifts to its customer base. Additionally, the development of more transparent tokenization metrics or standardized benchmarking will become increasingly important.