Cache billing adjustment
Prompt cache read tokens are now billed at an 85% discount off standard input token pricing — making long-context and repetitive workloads even more cost-effective.
Overview
Historical entry. This page records the policy introduced at that release. Current cache rates are model-specific; see Billing & pricing for the rates that apply now.
At the time of this release, cached prompt tokens — the portion of your input that the provider had already seen and could serve from cache — were standardized at 15% of the standard input rate. This reflected the upstream cost assumptions used for that release while still passing significant savings to users.
No changes to your API keys, request format, or model selection are required. Cache read tokens are automatically detected from provider usage reports.
How cache billing works
m_in).m_out).Why the adjustment
As prompt caching adoption has grown, the volume of cached tokens has increased significantly across the platform. This adjustment aligns cache read pricing more closely with the actual infrastructure cost of serving cached content, while preserving the core benefit: cached tokens remain far cheaper than full input tokens.
Impact
For most users, the change is modest. Workloads with heavy cache usage (long conversations, large system prompts, repeated context) will see the most noticeable adjustment, while requests with little or no caching are largely unaffected.
Existing credit balances and API keys are unchanged. The new rate applies automatically to all new requests.