Claude published a piece on 8 September 2026 arguing that cost and performance are not the trade-off most teams assume. The claim is specific: many applications running on Claude Platform can cut spend without giving up quality, using three levers. Cache hit rate, prompt instructions, and effort.
The prompt caching half of this is well covered already. What is newer, and what most Australian teams we work with have never touched, is the other two. Old instructions in your prompts are costing you money, and effort is a dial almost nobody calibrates.
How do you reduce Claude API costs without losing performance?
Three levers, in the order worth trying. First, raise the prompt cache hit rate so repeated context is read back cheaply instead of reprocessed. Second, strip instructions written for older models, because frontier models follow them literally and burn tokens doing so. Third, calibrate effort to the task, since a stronger model working lightly is often cheaper and better than a weaker model working hard. Claude's own reporting in September 2026 shows all three moving cost down while accuracy holds or improves.
Your old prompt instructions are the quiet cost
Prompts accumulate patches. Someone added a rule in 2024 because a model kept skipping a step, and the rule never came out. On a frontier model those patches are followed to the letter, which means extra reasoning, extra tool calls and extra tokens for no benefit. The published guidance names six patterns worth auditing:
Verification rituals such as double-check your work or verify twice before responding.
Thoroughness boosters such as be maximally thorough, or instructions shouted in capitals.
Mandatory procedures and scratchpad scaffolds that duplicate reasoning the model already does.
Stale few-shot examples tuned to an older model's failure modes.
Contradictory rules, which a more literal model will follow into a corner.
Dated configuration such as manual thinking budgets, which newer models can reject outright.
In a published test migrating a customer support workload from Opus 4.8 to Opus 5, removing planted anti-patterns of exactly this kind cut cost by 14.6% and raised accuracy by 5.3% on average across six prompts. Cost fell because duplicated reasoning and unnecessary tool calls stopped. Accuracy rose partly because a retired thinking setting had been causing the API to reject requests outright, and partly because contradictory refund rules had been making the model withhold refunds it should have issued.
Effort is the dial almost nobody turns
Effort tells Claude how hard to work. Low effort reaches a conclusion faster; high effort deliberates, verifies and explores alternatives. The published figures show the curve is neither flat nor linear, and that the top of it is frequently not worth paying for.
Published cost and score points across effort levels
| Model and effort | Benchmark | Score | Cost |
|---|---|---|---|
| Fable 5, low effort | FrontierCode Diamond | 11.5% | $5.35 per task |
| Fable 5, max effort | FrontierCode Diamond | 30.9% | $19.00 per task |
| Fable 5.1, low effort | Humanity's Last Exam | about 53% | about $0.30 per question |
| Fable 5.1, max effort | Humanity's Last Exam | about 61% | about $2.23 per question |
Read the last two rows carefully. The final step up to maximum effort adds about half a percentage point for roughly 46% more cost, and that gain sits inside the benchmark's run-to-run noise. You are paying more for nothing measurable. The reported CursorBench 3.2 result runs the other way and is more interesting: Fable 5.1 at low effort matched Fable 5 at high effort for about a third of the cost, helped by cache reads priced at $0.25 per million tokens against $1.00 for the older model.
Effort can be wrong in both directions. Set too high and the model deliberates past the point where there is any evidence left to find, adding cost and latency and sometimes degrading the answer. Set too low and it stops early, answers from the first search result, and produces something that looks finished but rests on partial information. Our note on choosing the right model and effort level in Claude Code covers the same dial from the developer side.
Caching, briefly
Cache reads cost a fraction of full input price, so the hit rate is the first thing to check. The short version: keep volatile values such as timestamps out of the prefix, do not reorder tool definitions, avoid changing effort mid-conversation on models where that breaks the prefix, and watch out for long tool calls that outlive the cache time-to-live. We have written the full treatment separately in prompt caching with the Claude API, so this piece will not restate it.
What the automated audits found
Claude also shipped commands that do this work for you: a prompt audit, a broader cost optimisation pass, and an iterative search that tunes model, effort and prompt against your own evaluation. Run across four public benchmarks starting from Sonnet 5, the cost optimisation pass reported reductions of roughly 58% on LegalBench, 73% on tau2-bench retail, 52% on OfficeQA Pro and 55% on SWE-bench Verified, with pass rates holding within noise.
The iterative search result is the one to show a finance director. Starting from a customer support setup on Opus 4.8 at default effort, it moved to Opus 5 at low effort with the anti-patterns removed, then down to Sonnet 5 at low effort with added routing rules. On held-out tickets the final configuration scored 90.5% against the original 78.6%, at about a fifth of the cost. Better and cheaper, from configuration alone.
Where we would start on an Australian workload
Order matters. Audit the prompts before you touch effort, because a prompt full of verification rituals will make every effort level look worse than it is. Then sweep effort across your own evaluation rather than trusting a benchmark. Then check caching. Most of the savings we find on client work are in the first two steps, and almost none of them need new code.
A cost review on a live Claude workload is typically a $6,000 to $12,000 piece of work for us, and on a team spending $8,000 a month it pays for itself inside a quarter. If you are running Claude in production and have never audited the prompts, that is the cheapest thing on your list. See our consulting services, or read Claude's own writeup for the engineering detail.
