Anthropic released Claude Opus 5 on 24 July at $5 per million input tokens and $25 per million output, with a reported 96.0% on SWE-bench Verified. The score is the headline. The more useful detail is that it arrived at half the price of the Fable line above it, which changes when reaching for the top of Anthropic's range is actually justified.
What the numbers say
Opus 5 carries a 1,000,000-token context window and a 128,000-token maximum output, taking text, images and files. Pricing is $5 and $25 per million tokens, with a fast mode at double the rate that runs roughly 2.5 times quicker.
On third-party aggregate measures it leads: the Artificial Analysis Intelligence Index places it at 61 and the Agentic Index at 55.3, ahead of Claude Fable 5 at 60 and 52.8, and ahead of GPT-5.6 Sol at 59 and 54.0. Anthropic reports 96.0% on SWE-bench Verified.
Those margins are narrow. One or two points on an aggregate index is not a difference you will feel on a given task, and it is well within the range where your particular workload could reverse the ordering entirely.
Extended thinking is now a dial you control
Opus 5 exposes a per-request reasoning effort setting with low, medium and high levels. Low is faster and cheaper for routine work; high lets the model think longer on genuinely hard problems.
This matters more than it sounds, because reasoning tokens are billed as output. A model that thinks for 2,000 tokens before writing a 300-token answer bills you for 2,300 tokens of output and hands you 300 tokens of value. Running everything at high effort is a quiet way to multiply your bill for work that did not need it.
The useful discipline is to default low, and raise the effort only for the categories of problem where you have checked that it changes the answer. On a lot of everyday work, it does not.
Where it sits against GPT-5.6 Sol
These two are the obvious comparison this month. Opus 5 is $5 and $25; GPT-5.6 Sol is $5 and $30 at launch. Near-identical input pricing, Opus slightly cheaper on output. The aggregate indices put Opus marginally ahead on intelligence and marginally behind on agentic work.
That is close enough that price and benchmark tables cannot decide it for you. The deciding factors are the unglamorous ones: which one handles your domain's vocabulary, which one follows your formatting instructions reliably, which one fails in ways you can detect.
None of that is on a leaderboard. It shows up when you put the same prompt in front of both and read the two answers next to each other, which takes a minute and settles the question better than any index.
Is the fast mode worth double?
Fast mode costs $10 and $50 for roughly 2.5 times the speed. Whether that is a good trade depends entirely on whether anyone is waiting.
For an interactive product where a user is watching a response arrive, halving the wait is often worth real money. For a background job, batch processing or anything where the result is consumed later, you are paying double for a property nobody experiences. The mistake we see most is enabling it globally because it sounded better.
Frequently Asked Questions
Should I use Opus 5 or Claude Fable 5?
Start with Opus 5. It scores higher on both published indices and costs half as much per token, so Fable now needs a specific reason rather than being the automatic choice for hard problems. If you have a workload where Fable measurably wins, keep it there, but verify rather than assume.
Is 96% on SWE-bench Verified as good as it sounds?
It is a strong number on a respected benchmark, and it is still a benchmark. SWE-bench Verified measures resolving real GitHub issues in specific repositories under specific conditions. Whether that predicts performance on your codebase, with your conventions and your test suite, is a separate question that only your own testing answers.
What does the reasoning effort setting cost me?
Reasoning tokens bill as output tokens. Higher effort means more thinking before the answer, which means a larger output bill for the same visible response. Treat effort as a cost dial, not a quality dial, and raise it only where you have confirmed it changes the result.
