Blog

Claude vs GPT-6 Prompt Caching: Hit Rate Beats Discount

November 2026 · 6 min read · Technical

Hand-drawn tap dripping into a bucket, with one drop missing the bucket
← Back to all posts

In September 2026 OpenAI announced better prompt caching for GPT-6: discounts of up to 90% on cached input tokens, a 30-minute window for reusing a cached prefix, and new tooling that includes a caching dashboard, a cache-miss diagnostics tool and explicit cache breakpoints. Claude has offered prompt caching with its own discount structure for some time, so a comparison table of the two discount rates is the obvious next step.

It is also the wrong place to start. The headline discount only applies to the share of requests that actually hit the cache, and that share is set by how your application is built, not by which vendor you use. The most telling part of OpenAI's announcement is that the early partner results are all about hit rates, not discount rates.

Is GPT-6 prompt caching cheaper than Claude prompt caching?

Not in any way you can read off a price sheet. Your effective saving is the cache discount multiplied by your cache hit rate. A 90% discount at a 50% hit rate saves less than a smaller discount at a 90% hit rate. Both vendors now offer deep discounts on cached input, so for most Australian businesses the bigger variable is whether their integration is structured to hit the cache at all, and whether anyone has measured it.

Here is what OpenAI reported from early partners. These are vendor-published figures, not independent measurements:

  • GitHub Copilot cut fresh-token processing by more than 50% across billions of requests.

  • Manus lifted its cache hit rate from roughly 85% to consistently above 90% in under a week.

  • Strawberry Browser moved from an 83% to a 91% hit rate and cut inference costs by 36%.

Look at what changed in the last two. The discount did not change; the hit rate did, by around six to eight points, and in Strawberry's case that was worth a third of the inference bill. The full detail is in OpenAI's prompt caching announcement.

Hit rate times discount: a worked comparison

Take an Australian business spending $20,000 a month on input tokens that are eligible for caching, such as a fixed system prompt, tool definitions and a reference document resent on every call. The table compares two hypothetical setups. The figures are illustrative, in AUD, and the discount rates are round numbers chosen to show the effect rather than either vendor's exact pricing.

Illustrative monthly input cost by cache discount and hit rate, AUD, as at November 2026
SetupCache discountHit rateMonthly input cost
No cachingNone0%$20,000
Deep discount, messy prompts90%60%$9,200
Smaller discount, tidy prompts80%90%$5,600
Deep discount, tidy prompts90%90%$3,800

The setup with the smaller discount and the tidy prompts costs $3,600 a month less than the one with the deeper discount and messy prompts. Moving from the second row to the fourth needs no change of vendor at all, only a change to how requests are assembled. That is the same pattern the Manus and Strawberry figures show.

What usually breaks the cache

Caching works on a repeated prefix, so anything that changes the start of a request on every call forces a miss. In the integrations we review, the usual causes are:

  • A timestamp, request ID or user name inserted near the top of the system prompt.

  • Tool definitions generated in a different order on each request.

  • Retrieved documents placed before the stable instructions instead of after them.

  • Chat histories that are trimmed or reshuffled from the front each turn, rather than appended to.

  • Traffic too sparse to keep the cache warm within the reuse window.

Most of these take minutes to fix once someone looks. The mechanics on the Claude side are covered in prompt caching with the Claude API and prompt caching lessons from Claude Code.

Where the real difference between vendors sits

If the discount rates are broadly comparable, what OpenAI has actually added is diagnostics: a dashboard to see the hit rate and a tool to explain individual misses. That is useful, and it is a fair challenge to everyone else. But a dashboard does not fix anything on its own. Someone still has to own the hit rate as a number, review it after every prompt change and treat a drop as a defect.

For a Claude deployment, you can get most of the same visibility today by logging the cache read and cache write token counts the API already returns on each response, and charting the ratio. It is a small piece of work. We often find it is the first time anyone in the business has seen the number.

Pricing structure matters too. A newer, cheaper cache rate does not always mean a cheaper task, as we found in our look at Claude Fable 5.1's cache pricing.

What not to conclude

Do not read OpenAI's partner numbers as a guide to your own savings. GitHub Copilot and Manus run very high volumes of near-identical requests, which is the best case for caching. A firm in Melbourne or Brisbane running a few thousand varied requests a day may never reach a 90% hit rate on either platform, and that is fine as long as it knows its actual figure. Equally, do not switch vendors to chase a cache discount. If your hit rate is low, it will be low after the migration too, and you will have paid for the move.

If you run Claude in production and nobody has checked the cache hit rate recently, that is usually the fastest cost win available. Our services include a cache review of existing builds, or contact us to have a look at yours.

FAQ

Frequently asked questions

How much is the GPT-6 prompt caching discount?

OpenAI says GPT-6 offers discounts of up to 90% on cached input tokens, with a 30-minute window for reusing a cached prefix. The saving you see depends on how often your requests hit the cache.

What is a good prompt cache hit rate?

High-volume products with repetitive requests report rates above 90%, but that is a best case. A lower rate can be fine for varied workloads, provided you measure it and fix avoidable misses such as changing prompt prefixes.

How do I check my Claude prompt cache hit rate?

Log the cache read and cache write token counts the Claude API returns on each response, then chart cached input tokens as a share of total input tokens over time and after each prompt change.

Should I switch AI providers for a better caching discount?

Rarely. If your cache hit rate is low because of how prompts are built, it will stay low on another platform. Fixing prompt structure usually saves more than moving vendors.

Ready to move from AI pilot to production?

We help mid-market Australian businesses deploy AI automations that actually reach production and deliver measurable ROI.