Redis LangCache pitches up to 90% LLM cost savings

Redis has just rolled out LangCache, bringing semantic caching to cut repeated AI-request costs. Addy Osmani says agents use about 4x the tokens of chat, citing Redis claims of up to 90% API savings. Critics question context drift, bad matches, and observability.

Redis LangCache pitches up to 90% LLM cost savings

TL;DR

  • Redis LangCache: Promoted as a managed semantic caching layer to cut repeated AI request costs
  • Token cost claim: Agents consume ~4× tokens versus chat; semantic caching cited as “up to 90%” API savings
  • Primary cost driver: Repeated questions framed as major production token spending source
  • Context sensitivity concerns: Risk of poor results when requests depend on changing or prior-turn context
  • Safety and observability: Questions on detecting incorrect semantic cache matches and measuring cache quality
  • Not persistent memory: Cost reduction contrasted with durable user modeling; framed as “agents with amnesia”

Addy Osmani is promoting Redis’s LangCache as a way to reduce the cost of repeated AI requests. In a post marked #ad, Osmani claims that agents consume roughly four times as many tokens as chat interactions and that semantic caching can reduce API costs by “up to 90%,” according to Redis.

The post presents repeated questions as a major source of token spending in production AI systems. LangCache is described as a managed layer for semantic caching, though the post does not establish whether the quoted savings apply broadly across different workloads.

Several replies questioned how caching would handle requests that depend on changing context. Dipankar Kumar Singh asked whether the approach could “break the prefix” when a model needs to receive previous context again. Mahmoud expressed concern that cached responses could produce poor results by losing relevant context.

Other replies focused on measuring the risks. Mikhail Rogov asked how teams could identify an incorrect semantic cache match, warning that “one wrong match can make that 90% saving expensive.” Punit Sethi similarly asked about observability for evaluating cache quality.

The discussion also distinguished lower token usage from persistent memory. One reply characterized the approach as “the bill version of agents with amnesia,” arguing that caching may reduce costs without giving an agent a lasting model of an individual user. Another commenter emphasized Osmani’s four-times estimate, suggesting that agents repeatedly process the same context across turns.

Source: Addy Osmani on X

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community