# OpenAI Adds 12-Hour Prompt Caching: Stop Truncating Agent Context

> OpenAI's 12-hour prompt caching removes the need to cut context to save costs. Anchor business rules in persistent prefixes for lower costs and stable runs.

Source: https://www.starboxai.cn/en/insights/ai-news/2026-10-21-openai-12h-prompt-caching-agents/ (Starbox AI Insights, updated 2026-10-21)

OpenAI's developer account announced on October 21, 2026, that API users can now select an optional 12-hour prompt cache retention window. Tibo at OpenAI also confirmed that logged-in users across Codex and Work retain persistent caches indefinitely, guaranteeing cache hits even when resuming old sessions hours later. 

This 12-hour update fundamentally resets the unit economics of enterprise AI agents. I am Xiao Xing from Starbox AI. Here is what this shift means for real-world business deployments.

A founder running an overseas precision instruments business recently shared a common headache: their customer support agent often deals with 3- to 4-hour gaps between replies due to time zones. To cut API spend, their engineering team set an aggressive timeout rule that wiped context after idle periods. Every time a customer replied, the agent re-read tens of thousands of tokens of product manuals, diagnostic workflows, and tool schemas. Not only did this fail to save money, but the agent frequently lost earlier context, repeatedly asking customers the same diagnostic questions and damaging user trust.

## Why 12-Hour and Persistent Caching Resets Agent Inference Costs

Under previous prompt caching setups, caches lasted only through short idle windows. If your workflow nodes ran several hours apart, the cache expired. The next call paid full price to process the entire input from scratch, driving up runtime inference costs.

With this update, long-running agents that process requests hours apart can reuse shared instructions, tools, and background context at reduced cached input rates. For logged-in users on Codex and Work, session caches stay active indefinitely.

Many technical teams operated under an outdated assumption: "shorter prompts mean cheaper calls." As a result, engineers aggressively trimmed prompts, truncated conversation histories, and chopped knowledge bases into fragmented snippets. In an era of short-lived caches, that approach shaved minor costs, but it broke context and forced models to guess intent. 

When cache lifetimes stretch to 12 hours or indefinitely, the winning strategy flips. You should front-load all immutable SOPs, API contracts, and brand knowledge into your system prompt once, letting the infrastructure lock them into cache memory.

## Implementing Long-Context Optimization: Rebuilding Your Persistent Prompt Foundation

With half-day cache windows, workflows must evolve from assembling small, dynamic prompts on the fly to managing structured, layered context. 

The core rule of long-context optimization is simple: cleanly separate your immutable foundation layer from your dynamic interaction layer. 

By pushing fixed content to the very beginning of the prompt, the model matches the cache from character zero.

| Prompt Layer | Contents | Change Frequency | Caching Strategy |
| :--- | :--- | :--- | :--- |
| **Layer 1: Business Foundation** | Brand definitions, industry glossaries, output constraints, compliance guardrails | Very Low (weekly or monthly) | Persistent / 12-hour cache hit |
| **Layer 2: Tool Contract** | External API JSON schemas, function definitions, data validation rules | Low (updated during system releases) | Placed directly after Layer 1, reused all day |
| **Layer 3: Task SOPs** | Decision trees, troubleshooting checklists, multi-step execution guides | Medium (updated with business process changes) | Reused in batches by task type |
| **Layer 4: Dynamic Session Zone** | User queries, runtime variables, execution logs | High (changes every turn) | Billed dynamically, excluded from static cache prefix |

## The Persistent Cache Migration Checklist

Use this checklist to audit and refactor your existing API implementations:

- [ ] **Check prefix token length**: Confirm that your static business base and tool schemas meet the minimum [Token](/en/glossary/token/) threshold required to trigger prompt caching.
- [ ] **Standardize static prefix order**: Audit every agent endpoint to ensure the first 80% of instructions, guidelines, and schemas are identical in order. Never place dynamic variables like timestamps or request IDs at the start of a prompt.
- [ ] **Separate tool schemas from runtime arguments**: Place tool definitions and schemas in the static prefix; append actual tool execution outputs at the end of the prompt.
- [ ] **Group workflows by 12-hour cycles**: Route workflows with long pauses—such as support tickets and [sales lead](/en/glossary/sales-lead/) follow-ups—into the 12-hour cache mode.
- [ ] **Monitor cache hit ratios**: Log the cache hit metrics returned by the API. If hit rates fall below targets, check whether dynamic code is injecting variables into your static prefix.

## What We See in Practice at Starbox AI

Across Starbox AI's work with 500+ enterprises, marketing and content teams frequently run into cost spikes and latency issues during multi-round editing cycles.

In our New Media Content Factory, writing agents draw from a centralized [brand knowledge base](/en/glossary/brand-knowledge-base/) and topic library, generating omni-channel content across nine messaging angles. Orchestrating these agents relies heavily on standardized, persistent knowledge. 

When foundational models extend prompt cache retention to 12 hours, systems managing large brand knowledge bases can sustain high-frequency, multi-turn agent collaboration at a fraction of the cost. Producing a piece of content drops from 2–4 hours down to 5–10 minutes not through blind trial and error, but because a structured, persistent knowledge base works continuously in the background.

A longer prompt cache looks like a pricing tweak on the surface, but it fundamentally shifts agent architecture. Fragmented workflows that rely on chopped-up prompts and aggressive context truncation are losing their edge. Teams with standardized operating procedures that turn deep business logic into persistent prompt foundations will pull ahead in efficiency.