Tech Stack
Claude Opus 5 shipped on July 24, 2026 as claude-opus-5. It costs $5 per million input tokens and $25 per million output — exactly half of Claude Fable 5's $10 and $50, and unchanged from Opus 4.8.
The headline everyone ran was the benchmark chart. The more useful story is that a cheaper model does not automatically produce a cheaper bill, and five defaults decide whether you actually capture the saving.
Everything below comes from Anthropic's published pricing docs and the Claude Opus 5 system card, including Table 8.1.A and the effort-level notes.
## 1. Switch your default — this is not a downgrade
The instinct with a cheaper model is to assume you are trading quality for cost. On agent work, that is not what the numbers say.
| Task type | Benchmark | Opus 5 | Fable 5 | |---|---|---|---| | Computer use | OSWorld 2.0 | 70.6 | 66.1 | | Business workflow automation | Zapier AutomationBench | 26.0 | 17.4 | | Agentic terminal / engineering | FrontierBench v0.1 | 43.3 | 33.7 | | Professional knowledge work | GDPval-AA v2 (ELO) | 1861 | 1747 |
For orchestration, tool-calling, computer use and workflow automation, Opus 5 is both cheaper and stronger. That combination is unusual enough to be worth acting on the same day.
Where Fable 5 still wins, and it matters if this is your workload: SWE-bench Pro (80 vs 79.2), FrontierCode 1.1 (53.5 vs 53.4), DeepSWE v1.1 (69.7 vs 68.8), and Humanity's Last Exam without tools (56.5 vs 56.3). These are narrow margins, but they cluster in deep single-repository code work. If that is your day, keep Fable 5 there and switch everything else.
And to kill a rumour: Fable 5 is not being retired. Anthropic's deprecation table lists it active with no end date before June 9, 2027.
## 2. Stop running max effort — this is the actual money
If you change one thing, change this.
On Zapier AutomationBench, Anthropic reports that at medium effort Opus 5 scored 24% at $0.89 per task — outperforming both Opus 4.8 and Fable 5 at less than half their cost.
Read that again, because it inverts the usual assumption. The cheap setting on the cheap model beat the expensive setting on the expensive model.
If medium is not enough for your task, the next step is not max. It is xhigh, which on GDPval-AA still outperforms every other model while using 25% fewer output tokens than max. On AA-Briefcase, high effort beats every other model at 33% fewer tokens.
The pattern is consistent: the top of the effort range buys you a small amount of quality for a large amount of money. Most production workloads should live at medium, escalating deliberately rather than by default.
Practical version: set medium as your baseline, log which task types actually fail there, and raise effort only for those. That single change is usually a bigger cost lever than switching model.
## 3. Delete your carried-over verification instructions
If you have prompts that say *"double-check your work"*, *"verify each step before continuing"*, or *"review your answer for errors"* — instructions that were genuinely useful on older models — Anthropic now advises removing them.
Opus 5 self-verifies. Those instructions cause it to over-verify, which means you pay twice for the same check, in tokens and in latency.
This is the rare optimisation that takes two minutes, requires no testing infrastructure, and makes the output better rather than worse. Grep your prompt library for "verify", "double-check" and "make sure you" and delete what the model now does on its own.
## 4. Add tools, not thinking budget
This is the most useful line in the entire release, and it is buried in the system card:
> Agentic tool-use is generally a more cost-effective method of scaling test-time compute than adaptive thinking by itself.
That is an architecture instruction disguised as a footnote.
When an agent underperforms, the reflex is to raise the thinking budget — let it reason longer. Anthropic's own evaluation says the cheaper and more effective move is usually to give it a better tool: a calculator instead of mental arithmetic, a search instead of recall, a validator instead of self-inspection, a real API instead of a guess.
Thinking scales cost linearly and quality logarithmically. A correct tool converts a hard reasoning problem into a cheap lookup. If you maintain agents, this reorders your backlog.
## 5. Fix the two things that will break silently
Thinking is now on by default. Sending thinking: {"type": "disabled"} returns a 400 error at the xhigh and max effort levels. If you have workloads that explicitly disabled thinking on Opus 4.8, they will fail outright — or worse, behave unexpectedly.
There is a subtler version of the same trap. max_tokens caps thinking and response together. A workload previously tuned with thinking off, using a tight max_tokens, can now spend that budget on reasoning and silently truncate the answer. Nothing errors. The output is just quietly incomplete.
Audit anything with an aggressive max_tokens before you switch.
## The free saving nobody will notice
The minimum cacheable prompt dropped from 1,024 tokens to 512.
If your system prompt sat between those numbers, it could never be cached — and you were paying full input price for identical text on every single call. It now caches with no code change at all. For a high-volume endpoint with a 600-token system prompt, that is a real line-item reduction for zero work.
Cache reads on Opus 5 are $0.50 per million against Fable 5's $1.00, so the saving compounds.
## One caution before anyone celebrates
Opus 5 fast mode is $10 in and $50 out per million tokens. That is *exactly* Claude Fable 5's standard rate.
The half-price advantage exists only at standard speed. If you switch to fast mode assuming you are still saving money, you have quietly moved back to flagship pricing. Fast mode may well be worth it when latency is the constraint — just do not confuse it with the cheap option.
Also note: setting inference_geo: "us" applies a 1.1× multiplier across every pricing category. And the 1M context window is included at standard pricing with no long-context premium, which is a genuine improvement over models that penalise long inputs.
## What I actually changed
For the record, my own setup after reading the card:
- -Opus 5 as the default for orchestration, tool-calling and content work
- -Fable 5 retained only for deep single-repo code work
- -Medium effort as the baseline, escalating to xhigh by exception and never to max by default
- -Every "verify your work" line removed from my prompt library
- -One agent's thinking budget cut and replaced with a proper validation tool
- -Audited
max_tokenson everything that previously ran thinking-off
## The takeaway
Cheaper models do not save you money. Correct settings do.
A team that switches to Opus 5 and leaves max effort on, keeps redundant verification instructions, and never caches its system prompt can comfortably spend more than it did on the pricier model. The saving is real, but it is in the configuration, not the price list.
Every figure in this piece comes from Anthropic's published pricing documentation and the Claude Opus 5 system card. Where two Anthropic sources describe the same result differently, I have used the system card, since it carries the numeric table.
The same logic one layer down: cost follows architecture, not price lists. Cloud inference bills per use, so it has to be rented; local inference bills once, so it can be owned. I applied that to my own dictation tool — the transcription runs on your Mac, which is why [it can charge once](https://indianwhisper.com/pricing?src=blog-claude-opus-5-settings-guide) rather than monthly.
Want to build something like this?
I architect and deploy end-to-end AI systems — from MVP to revenue.
Let's TalkOr ask Angelina — my AI twin in the bottom-right corner. She knows my full build history, live GitHub, and how I'd approach your project.