Anonymous request-token comparisons from Opus 4.6 and Opus 4.7 (tokens.billchambers.me)

Article summary

The article appears to discuss the differences between Opus 4.6 and Opus 4.7, with a focus on the tokenizer change and its potential impact. The discussion reveals that the new tokenizer can output more tokens for the same input, leading to increased costs. The article's content is not directly available, but the comments suggest that the change has significant implications for users. The comments also touch on the potential consequences of this change, including increased costs and decreased performance.

Main themes

• Opus 4.6 vs 4.7
• Tokenizer changes
• Cost and performance implications
• AI model development
• Pricing and accessibility

What commenters say

• The new tokenizer in Opus 4.7 can output more tokens for the same input, leading to increased costs and potentially decreased performance.
• Some users believe that the changes in Opus 4.7 have resulted in a less capable model, with one user suggesting that it has been 'lobotomized'.
• Others argue that the increased cost of using Opus 4.7 is justified by its improved capabilities, and that the cost will decrease over time as training costs fall.
• There is concern that the pricing model for AI tools is becoming increasingly expensive and inaccessible to individual users, with some suggesting that this is a deliberate effort to raise the baseline for AI engineering tasks.
• Some commenters speculate that the high cost of AI tools is due to the high training costs, and that companies need to charge high margins to recoup their investments.
• Others believe that the value of AI models will decrease over time as they become commoditized, and that the only moat for companies will be their data.
• There is also discussion about the potential for open-source or local models to be released, which could give proprietary companies a run for their money.
• Some users are frustrated with the changes in Opus 4.7, citing decreased performance and increased frustration when interacting with the model.