#promptcaching

GPT-5.6 added explicit prompt cache controls and quietly took the 24-hour cache away. We ran the API and measured what that swap costs.

A numbers-first 2026 guide to cutting LLM API bills: how prompt caching, batch APIs, and model routing stack across Anthropic, OpenAI, and Gemini.

Eviction, offload-and-recall, retrieval-over-stuffing, and subagent isolation — the four ways to keep an LLM coding agent from drowning in its own context.

Your Claude Code cache hit rate is probably below 50%. Anthropic's own team treats sub-90% as a SEV. Here are the 14 cache-break patterns and 6 rules that cut m

OpenAI now keeps prompt caches for 24h by default on GPT-5.5. We ran the API to see when the 90% discount actually shows up, and when it doesn't.

I ran five reviewer agents over a 16K-token PR. The naive setup cost $1.32. Two architectural...

How Anthropic's prompt cache works mechanically — the ephemeral cache_control marker, the two-tier write premium (1.25x for 5-min TTL, 2x fo

OpenAI's prompt cache engages automatically on prompts ≥1,024 tokens with no caller-side configuration. The mechanics, the 90% discount

시스템 프롬프트·페르소나·문서가 호출마다 동일하다면 그 부분을 캐시할 수 있습니다. Anthropic·OpenAI의 prompt caching 작동 원리와 마케팅 자동화의 비용을 90% 깎는 운영 패턴.

Real first-party data: Anthropic prompt caching cut Citare's AI bill 25-35% on parsing-heavy workloads. What works, what doesn't, what burned me $20.

Continuing the cache TTL audit series, this time breaking down by model. The server silently switched main agent models three times (Opus 4.6 → Sonnet 4.6 → Opus 4.7 → Sonnet 4.6), and Claude Code autonomously assigns sub-agent models. Comparing cost per million main output tokens across 7 periods reveals an 11.5x efficiency gap.

Сократить расходы на LLM API позволяют пять приёмов, которые работают независимо от провайдера и...