OpenAI shipped a Prompt Caching Dashboard for GPT-6, and the new models keep the cache warm when an agent changes reasoning effort or its tool list

An agent that reaches for a heavier reasoning setting on a hard step used to lose everything it had already cached. Meredith Shubel at The New Stack reports OpenAI released a separate prompt-caching announcement alongside the GPT-6 Sol and Luna model launch on 22 September 2026, and the biggest change is that agents can now vary the reasoning effort and the list of tools they pass in without breaking the cache on the earlier context.

The New Stack quotes OpenAI's own statement: "We've improved prompt caching for GPT-6 to deliver higher cache hit rates by default, helping agents reuse more context, respond faster, and benefit from discounts of 90% on cached input-token reads." The 90% discount rate itself matches GPT-5.6; the change is how often the cache actually gets hit.

The dashboard makes cache reuse a number a team can read

OpenAI also shipped a Prompt Caching Dashboard. The New Stack reports developers can view how much input is cached, how that amount changes over time, and the diagnostics tool flags missed caching opportunities so a team can see which parts of a prompt or agent loop are being re-processed on every call. That is a measurement any team paying for GPT-6 tokens can now run on its own traffic instead of guessing.

The other change is what OpenAI calls preserving earlier context when reasoning effort and tool availability change. The New Stack describes it as an agent scaling reasoning effort up and down based on how difficult a step is, then making different tools available depending on what the task requires, without disturbing earlier cached context. In the previous model families, changing either setting mid-run tended to invalidate the cache, so the same prefix had to be processed again from scratch.

The 50 percent number is GitHub's, quoted by OpenAI, over several months

OpenAI says the improvements were already reaching GPT-5.6 users. The New Stack quotes OpenAI attributing to GitHub the claim that "these improvements have reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models," measured across the past "several months." That is OpenAI reporting GitHub's number, and neither figure comes from an independent benchmark, so treat it as a vendor claim.

The New Stack reports the API prices for Sol and Luna are down 50% against their GPT-5.6 counterparts, and 58% lower on Luna output tokens. Cheaper tokens and a higher cache hit rate are two separate savings, and both apply to the same requests. A team building on GPT-6 has one measurement it can make today with no code change: open the Prompt Caching Dashboard and read the real cache hit rate on the last week of traffic. If it is close to what OpenAI advertises, the pricing math for the agent has already shifted. If it is well below, the dashboard says which parts of the prompt or loop are the reason.

Source

Meredith Shubel, The New Stack, OpenAI cut GPT-6 token prices in half. The bigger lever may be the cache., 23 September 2026. Vendor announcement: OpenAI, Better prompt caching for GPT-6, 22 September 2026.