Anthropic API / Anthropic API

Prompt caching not working on Opus 4.5 and Haiku 4.5 (works correctly on Sonnet 4.5)

Fix prompt caching not being applied on Claude Opus 4.5 and Haiku 4.5 models despite cache_control parameter set Includes evidence for Anthropic API troubleshooting demand.

Source-backed Last updated May 21, 2026 1 source Needs local verification

Category: Anthropic API
Error signature: prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected
Quick fix: Compare the failing environment with a known working setup, then change one configuration value at a time.
Updated: May 21, 2026
Verification status: Source-backed
Evidence: 1 public source URL

Before you change production

This page includes public source URLs in the imported troubleshooting record. Compare those references with your version and environment before applying changes.

Reproduce the smallest failing action and save non-secret logs before changing configuration.
Check versions for Anthropic API, related SDKs, package managers, CI runners, and hosting providers.
Change one setting or dependency at a time, then rerun the same failing command or request.
Avoid destructive commands, credential rotation, billing changes, or security relaxations without a rollback plan.

What this error means

prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected is a Anthropic API failure pattern reported for developers trying to fix prompt caching not being applied on claude opus 4.5 and haiku 4.5 models despite cache_control parameter set. Based on the imported evidence, treat this as a tool-specific troubleshooting page rather than a generic API error.

Why this happens

GitHub issue anthropics/anthropic-sdk-python#1085 — 7 comments, open since Nov 2025. Affects billing cost directly because uncached prompts consume full token count. New model-specific regression. Not in covered-errors.md. Category: Anthropic API — cache behavior bug affecting production costs.

Common causes

GitHub issue anthropics/anthropic-sdk-python#1085 — 7 comments, open since Nov 2025. Affects billing cost directly because uncached prompts consume full token count. New model-specific regression. Not in covered-errors.md. Category: Anthropic API — cache behavior bug affecting production costs.

Quick fixes

Confirm the exact error signature matches prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected.
Check the Anthropic API account, local tool state, and provider configuration involved in the failing workflow.
Compare the failing environment with a known working setup, then change one configuration value at a time.

Platform/tool-specific checks

Verify the command, editor, extension, or API client that produced the error.
Compare local settings with CI, deployment, or editor-level settings when the error appears in only one environment.
Avoid deleting credentials, local model data, or project settings until the failing scope is clear.

Step-by-step troubleshooting

Capture the exact error message and the command, editor action, or request that triggered it.
Check whether the failure is account/auth, quota/rate, model/provider, local runtime, or deployment configuration.
Review the source evidence below and compare it with your environment.
Apply one change at a time and rerun the smallest failing action.
Keep the working fix documented for the team or deployment environment.

How to prevent it

Keep provider/tool configuration documented.
Record non-secret diagnostics such as tool version, provider name, model name, and command path.
Add a lightweight check before CI or production workflows depend on the tool.

Diagnostic flow for this page

Match prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected exactly before applying the quick fix.
Compare the failing environment with Anthropic API versions, account scope, provider settings, and deployment context.
Check the listed common causes in order, starting with the cause that best matches your logs.
Use the evidence status below to decide whether to confirm against public sources or official documentation.
Apply one reversible change, rerun the smallest failing action, and keep rollback notes.

FAQ

What does prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected mean in Anthropic API?

prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected indicates the workflow described by "Prompt caching not working on Opus 4.5 and Haiku 4.5 (works correctly on Sonnet 4.5)" failed in a Anthropic API context. Start by confirming the exact signature and the component that emitted it.

What should I check first for prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected?

Check whether this cause matches your environment: GitHub issue anthropics/anthropic-sdk-python#1085 — 7 comments, open since Nov 2025. Affects billing cost directly because uncached prompts consume full token count. New model-specific regression. Not in covered-errors.md. Category: Anthropic API — cache behavior bug affecting production costs.

Is the prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected guidance source-backed?

The "Prompt caching not working on Opus 4.5 and Haiku 4.5 (works correctly on Sonnet 4.5)" page is labeled source-backed. It has 1 public source URL from github.com, so use that evidence status when deciding how much additional verification you need.

When should I avoid the quick fix for prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected?

For "Prompt caching not working on Opus 4.5 and Haiku 4.5 (works correctly on Sonnet 4.5)", avoid applying "Compare the failing environment with a known working setup, then change one configuration value at a time." when the error text differs, the affected version or provider is different, or the change would touch credentials, billing, DNS, production deploy settings, or irreversible data.

What should I collect before debugging prompt_caching not activated on new model families (Opus 4.5/Haiku 4.5) — Cache control header ignored, costing more tokens than expected?

For "Prompt caching not working on Opus 4.5 and Haiku 4.5 (works correctly on Sonnet 4.5)", collect the failing Anthropic API command or request, non-secret configuration, versions, timestamps, request IDs when available, and the smallest reproduction that still triggers the signature.