Anthropic downgraded cache TTL on March 6th (github.com)

Article summary

Anthropic's cache TTL was changed from 1 hour to 5 minutes around March 6th, causing a significant increase in cache creation costs and quota consumption for subscription users. The change was not announced and appears to have been made server-side. The author of the issue analyzed session data and found that the 1-hour TTL was the intended default, and the reversion to 5-minute TTL has caused a 20-32% increase in cache creation costs. The author requests that Anthropic confirm or deny the change and consider restoring the 1-hour TTL as the default.

Main themes

• Cache TTL change
• Cost and quota impact
• Anthropic's infrastructure
• LLM performance
• Vendor lock-in
• AI model degradation

What commenters say

• The cache TTL change may be related to Anthropic's peak-hour announcement and infrastructure load.
• The change has caused significant financial losses for users and has led to a loss of trust in the platform.
• Some users have switched to alternative AI models, such as Codex, due to the issues with Anthropic's platform.
• The output of AI models may be starting to get back into the training input, causing model collapse.
• The leadership of AI companies, including Anthropic and OpenAI, has been criticized for their business practices and ethics.
• Some users believe that the AI models are being intentionally throttled to increase revenue, while others think it may be due to infrastructure constraints.
• The use of AI models has led to a loss of coding and debugging skills, making users reliant on the models.
• The degradation of AI models may be due to the rapid growth in usage and the inability to scale compute quickly enough.