Claude Opus 5 launched on Friday July 24 with a half-price line attached. Anthropic says it comes close to Fable 5 at half the cost, and by Monday an independent test had measured the whole task bill at about 20% less. Claude Opus 5 is a good model sold with a flattering comparison, and the gap between those two numbers is the story.
Key points
- What does Claude Opus 5 cost per token? Claude Opus 5 lists at $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8, according to Anthropic’s announcement .
- Why is the measured saving 20% and not 50%? Because a rate card sets the price of each token, and the bill depends on how many tokens a task burns.
- What does the effort dial do to the bill? It moves the cost more than the model choice does.
- How should a small firm test the effort setting? Run the same 20 real tasks at high and at maximum, score the outputs blind, and put the two bills side by side with whatever usage data your provider shows.
What does Claude Opus 5 cost per token?
Claude Opus 5 lists at $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8, according to Anthropic’s announcement. Anthropic pitches it as near-Fable intelligence at half the price. The rate card is the easy part to check, and it matches the old Opus price, which Anthropic left at the Opus 4.8 level.
The claims about cost per task are harder to check. On CursorBench at maximum effort, Anthropic says Opus 5 lands within 0.5% of Fable 5’s peak score at half the cost per task. On OSWorld 2.0 it says Opus 5 beats Fable 5’s best result at just over a third of the cost. Those are internal results, and the post doesn’t print the cost figures behind them. For context, OpenAI launched GPT-5.6 Sol at $5 in and $30 out per million tokens earlier in July, so Opus 5 undercuts it on output rates. The list price held steady last time too, and Claude Opus 4.8 pricing shows how a bill can jump anyway.
Why is the measured saving 20% and not 50%?
Because a rate card sets the price of each token, and the bill depends on how many tokens a task burns. Artificial Analysis measured the whole task and got $17.79 for Opus 5 at maximum effort against $22.30 for Fable 5, about 20% less.
Opus 5 took 103 turns per task at that setting, against 55 for Opus 4.8. More turns means more tokens, and that eats into a lower per-token price.
It’s like a car sold on its fuel economy figure when what you pay depends on how far it has to drive to do the job. Newer models often think longer. A cheaper rate card paired with a longer thought process can land you a modest saving, or in the wrong workload a higher bill. Neither number tells you what your own tasks will do, which is why a trial on your real work stays the only reliable measure, as we said about Kimi K3’s benchmark table ten days earlier. The same logic sits behind OpenAI’s cost per successful task idea and our piece on GPT-5.5’s price per token.
What does the effort dial do to the bill?
It moves the cost more than the model choice does. Artificial Analysis measured Opus 5 at three effort settings. The cheapest came in at $10.41 a task against $17.79 for the dearest, with turns falling from 103 to 76. The table shows what each setting cost and what it scored.
| Effort setting | Cost per task | Turns | AA-Briefcase score |
|---|---|---|---|
| Maximum | $17.79 | 103 | 1,720 |
| Extra-high | $14.26 | 91 | 1,693 |
| High | $10.41 | 76 | 1,606 |
For comparison, Fable 5 costs $22.30 a task and scores 1,574 on the same index, with GPT-5.6 Sol at 1,505. The article doesn’t say which setting Fable 5 ran at. High effort at $10.41 is about 53% below Fable 5’s cost and still ahead of it on that index. The price of dropping from maximum to high is real, about 114 points, and it’s one benchmark.
Our view is that high effort is the right default for routine work, with maximum reserved for tasks where a wrong answer is expensive. Dials are free to turn. A model swap isn’t.
How should a small firm test the effort setting?
Run the same 20 real tasks at high and at maximum, score the outputs blind, and put the two bills side by side with whatever usage data your provider shows. You’re looking for the tasks where maximum earned its price.
Our guess, and it is only a guess, is that most routine work shows no difference you’d pay for, and a few hard tasks show a clear one.
If you find that split, you don’t need one setting for the whole firm. Default to high, and write down which task types get promoted to maximum. That list is the real output of the trial. It tells your staff when to ask for more thinking, and it gives you a figure to put against the next price change.
One more thing to watch. Turn counts are what moved the bill here, and a model that takes 103 turns where its predecessor took 55 is working differently as well as charging differently. If a vendor won’t tell you how many steps a task took, you can’t see the real price.
What would 500 tasks a month cost?
Here’s an illustration, not a forecast. If your tasks cost what Artificial Analysis’s did and you ran 500 a month, Fable 5 would run $11,150, Opus 5 at maximum $8,895 and Opus 5 at high effort $5,205. The first swap saves you about a fifth. The second setting change saves you more than the swap did.
The sceptic’s answer is that a 20% saving on one benchmark is a rounding error once a model changes how your staff work, and that launch-week prices rarely survive a quarter of real use. Both points deserve respect, and a saving that small might fall short of a reason to switch. But the 20% is the maximum-effort number. The same test shows the dial can take the saving past half.
Read the headlines like a buyer. Half the price is a claim about rates. About 20% less is a measurement of a bill. About 53% less at high effort is a measurement with a quality cost attached. Three different claims share one product, and only the last two come with units you could check.
Check what your own tasks cost before you believe any of it. Our guide to AI costs for Canadian business covers the wider picture, and our comparison of ChatGPT, Claude and Gemini for business covers the choice between vendors.
We haven’t run Opus 5. The figures come from one index built for agentic knowledge work, Anthropic’s half-price and one-third-price claims are internal and unaudited, and a cost per task on one benchmark can’t be carried over to your invoices. Treat everything here as dated July 27. If Anthropic prints its own cost-per-task data and it matches the independent figure, we’ll say so. We’d also change our view if your tasks needed far fewer turns than the benchmark’s, because then the per-token price would matter more than we’ve said.
Start at high because the cost of being wrong about it is small and visible. If a few answers come back weaker, you promote those task types and carry on. The cost of being wrong about maximum is a bill that runs about 70% higher per task on the same work, and nobody notices until the month closes. Start low, promote on evidence, and keep the list of promoted tasks short enough to fit on one page.
Run your 20 tasks at high effort first. Move up a setting only where a wrong answer costs more than the difference.
Anthropic has the harder job here. It has to show its claim holds on the work you do, and so far only an outside tester has put a task bill on the page.
Frequently asked questions
How much does Claude Opus 5 cost?
Anthropic lists $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. Your actual bill depends on how many tokens each task uses.
Is Claude Opus 5 half the price of Fable 5?
Per token Anthropic frames it that way. Artificial Analysis measured about 20% less per task at maximum effort, because Opus 5 used more turns.
What does the effort setting change?
It changes how long the model thinks, and so how much a task costs. Artificial Analysis measured $17.79, $14.26 and $10.41 a task at maximum, extra-high and high effort.
Claude is made by Anthropic, and we have not tested the model. Archive entry dated July 27, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
For the full method, see our column on how to measure cost per successful task before comparing model prices.
