Most teams reach for a cheaper model when the bill arrives, and most of them are leaving 60% or more on the table first. Four levers cut Claude Opus 5 spend substantially without touching output quality.
1. Prompt caching (the big one)
If any part of your prompt repeats — a system prompt, a document corpus, a tool list, a knowledge base — cache it. Cached reads cost roughly a tenth of the standard input rate.
Two Opus 5 specifics:
- The minimum cacheable prefix is 512 tokens, down from 1,024 on Opus 4.8. Prompts you previously wrote off as too short now cache with no code change.
- Caching is a prefix match. One changed byte anywhere in the prefix invalidates everything after it. A timestamp or a session ID interpolated into your system prompt silently destroys your hit rate — check
cache_read_input_tokensin the response to confirm it is actually working.
2. Batch anything that can wait
The Batch API costs 50% less on both input and output: $2.50 and $12.50 per million tokens. Most batches finish within the hour. If a job is not blocking a human, it belongs here — content generation, bulk classification, nightly analysis, evaluation runs.
3. Tune effort per route
The default effort level is high, which is not right for everything. Opus 5 is unusually strong at low and medium, so sweep your own evaluation set rather than assuming.
Measure cost per completed task, not cost per request. On agentic work higher effort often means fewer turns and less rework, so cranking effort down can cost more overall.
4. Shorten the output
Output tokens cost five times input tokens, and Opus 5 writes long by default. A short conciseness instruction cuts user-facing response length by roughly a fifth; an explicit word count cuts it further. Files the model writes to disk — reports, documentation — run longer still and need their own length instruction.
Note that lowering the effort level does not reliably shorten output. Prompting is the lever here.
5. Route instead of downgrading
The biggest saving is structural: run Sonnet 5 or Haiku 4.5 as the default and escalate to Opus 5 only when a check fails. Because most workloads have a long tail of easy requests, the blended cost lands far closer to the cheap model than the expensive one — while the hard cases still get the best model you have.
More on Claude Opus 5
Start with the complete Claude Opus 5 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Claude Opus 5 guideFrequently Asked Questions
How much can prompt caching save on Claude Opus 5?
Cached input reads cost roughly a tenth of the standard rate, so a workload dominated by a repeated prefix — a large system prompt, a document corpus, a knowledge base — can see input costs fall by most of their value. On Opus 5 the minimum cacheable prefix is 512 tokens, half what Opus 4.8 required.
Why is my prompt cache not working?
Almost always a silent invalidator in the prefix. Caching is an exact prefix match, so a timestamp, a UUID, a session ID, or non-deterministic JSON key ordering near the start of your prompt invalidates everything after it. Check `cache_read_input_tokens` in the response — if it stays at zero across identical requests, something upstream is changing.
Is it cheaper to use a smaller model or optimise Opus 5?
Do both, in that order: optimise first, then route. Caching and batching alone can cut spend dramatically without any quality change, and routing routine work to Sonnet 5 or Haiku 4.5 while escalating hard cases to Opus 5 usually beats picking one model for everything.
