Share GPT-6 aims to enable persistent agents to work for hours on complex tasks such as reorganizing codebases, preparing researched documents and presentations. In these agents’ successive API requests, repeated instructions, tool definitions and context are cached, allowing computation to be reused; response times are reduced, and developers can receive discounts of up to 90% on cached input tokens. The GPT-6 family introduces a new caching system that offers higher hit rates by default. Eligible common prefixes qualify for discounts when reused within a 30-minute window.
- The Prompt Caching Dashboard shows the amount of input served from the cache and hit rates over time; it can be used to compare cached and uncached tokens.
- The prompt caching diagnostics tool compares changes in models, tools, settings or inputs when cache misses occur. In the example, the reason is given as “tools_changed”, the comparable reusable token count as 5629, and the missed token count as 5629.
- Explicit cache breakpoints make it possible to choose which instruction prefixes will be stored.
- In GPT-6 models, reasoning effort can be changed between responses without breaking the cache.
- Tool definitions, schemas and ordering can be kept fixed; allowed_tools or tool_choice can be used.
- New developer messages can be appended to the end of the context to override previous instructions.
- Prewarming prepares shared context before the user request. Codex, meanwhile, helps review code and apply improvements.
Why it matters
This update makes it easier to manage the cost and latency caused by long-running agents reprocessing the same instructions, tool information and context at every step. Developers of applications that work particularly with codebases, generate documents or make multiple API calls will be able to assess performance not only by the model’s response, but also by how effectively the cache is used. The dashboard and diagnostic tool make the debugging process more concrete by showing whether a cache miss was caused by changes to tools, settings or inputs. The ability to use explicit breakpoints and change the reasoning level establishes a more controlled relationship between agent behavior and cache usage. However, the extent of the gains remains dependent on how well shared prefixes and tool configuration can be preserved throughout the workflow.
Background
GPT is not a new name in the FikirPilot archive: we have published 4 news articles mentioning the name in the last 90 days; the most recent one is dated September 29, 2026.
Term: agent
An artificial intelligence agent is software that calls tools and carries out multi-step tasks to achieve a goal instead of generating a single response.