Azure will start billing your prompt cache writes

Prompt caching on Azure has been half-free. Cache reads bill at a discount on input tokens; cache writes have cost nothing. That ends. Microsoft's Azure OpenAI pricing page now carries a notice that cache write charges are not active yet and that billing is "expected to start on or after August 21, 2026." That date is today. The rate column still reads $-. If you run agents on GPT-5.6 or later in Azure, you have a new line item with no published price attached to it.

What actually happened

Two Microsoft documents tell the story together. The Azure OpenAI pricing page says cache write charges will not appear on your bill at this time, with billing expected on or after August 21. The prompt caching how-to on Microsoft Learn, updated August 12, spells out the mechanism: models before the GPT-5.6 family do not charge extra to write to the cache, and on GPT-5.6 and later families, cache writes can incur charges on top of discounted cache reads.

The plumbing shipped ahead of the billing. On Standard pay-as-you-go deployments with GPT-5.6 and later, the usage object now reports cache_write_tokens alongside cached_tokens. You get prompt_cache_key to group requests that share a prefix, and explicit prompt_cache_breakpoint markers to say where a reusable prefix ends. Each request can create at most four new cache writes. In the default implicit mode, the breakpoint on the latest message eats one of those four slots automatically.

Microsoft has not published the multiplier. On the direct OpenAI API, cache writes bill above the base input rate — do not assume Azure matches it. Check the pricing page for your region before you model anything.

Why prompt cache write billing matters for your business

Agent loops are write-heavy by construction. A chat turn appends to the context and writes a new cache entry. A ten-turn agent session with a long system prompt and retrieved documents can write the cache on every turn. Under the old pricing that was free instrumentation. Under the new pricing it is a bill that scales with conversation length, not with useful work.

Your prefix discipline is now a cost control. Caching needs a minimum of 1,024 tokens and an identical first 1,024 tokens. One changed character — a timestamp, a session ID, a shuffled tool list — is a miss and a rewrite. Move stable content to the front, keep dynamic content at the end, keep conversation context append-only. That was always good practice. Now it has a price tag.

There is a kill switch, and you should know where it is. Set prompt_cache_options.mode to explicit and supply no breakpoints, and the request uses no prompt caching and incurs no cache-write charges. For short prompts that never hit the 1,024-token floor anyway, that is the honest configuration.

PTU-M users are in a different boat. Provisioned Throughput managed deployments still cache, but they do not support breakpoints and do not expose cache_write_tokens. If you are on PTU-M, you cannot measure this the same way — which is its own argument for logging usage objects across every deployment type you run.

Do the measurement this week, not after the first surprise invoice. Log cache_write_tokens and cached_tokens per request, and compute the ratio of writes to subsequent reads. A prefix you write ten times and read once was never a cache. It was a tax you had not been charged for yet.

Sources: Azure OpenAI Service pricing, Microsoft Learn: Prompt caching with Azure OpenAI.