Claude Opus 5: What Makes It a Leap, and How to Get the Best Out of It
Claude Opus 5 tops the main independent intelligence ranking and doubles its predecessor on agentic coding, at half the price of Fable 5. The real benchmarks, why the praise is earned, and how to prompt it well.
On this page
Short answer: On July 24, 2026, Anthropic shipped Claude Opus 5, and for once the reaction matches the numbers. It sits at the top of the main independent intelligence ranking, it roughly doubles its own predecessor on agentic coding, and it does all of this at the same price as Opus 4.8, which is half the price of Anthropic's flagship, Fable 5. I build AI features into production apps, so here is what actually changed, why the praise has been this loud, and how to get the best out of it.
What actually changed in Opus 5
Opus 5 is the successor to Opus 4.8, and the headline is that it closes most of the gap to Anthropic's frontier model while costing half as much per token. Pricing is $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. It has a one million token context window and can return up to 128,000 tokens in a single response.
The clearest place to see the jump is agentic coding. On Frontier-Bench v0.1, a benchmark that scores an agent doing multi-file changes, debugging, and building features from a spec in a terminal, Anthropic reports that Opus 5 "more than doubles Opus 4.8's performance at a lower cost."
Look closely, because there is a bigger story in that chart than "new model beats old model." Opus 5 also beats Fable 5, the more expensive flagship, on this benchmark.
The second number that made people sit up is ARC-AGI 3, a test built specifically to resist memorization and reward genuine novel reasoning. Anthropic reports Opus 5 scoring "three times as high as the next-best model."
A move from 1.5 to 30.2 in a single model generation is the kind of jump that does not usually happen, and it is on the one benchmark designed to be hard to game.
Two more advances matter if you are building real products rather than reading a scoreboard:
- Computer use. On OSWorld 2.0, which measures an agent driving an actual desktop, Anthropic says Opus 5 surpasses Fable 5's best result at just over a third of the cost.
- Science. Opus 5 scores 10.2 percentage points higher than Opus 4.8 on organic chemistry, and 7.7 points higher on protein tasks.
Why is Opus 5 getting so much praise?
Three reasons, and none of them are marketing copy.
First, an independent scoreboard rather than Anthropic's own. Artificial Analysis runs its own evaluations instead of repeating vendor claims, and it places Opus 5 at 61 on its Intelligence Index, ranked first out of the 189 models it tracks. The median model on that index scores 33, so this is not a crowded photo finish.
Second, the price. Opus 5 lands within 0.5% of Fable 5's peak score on CursorBench 3.2, at half the cost per task. For most of the last two years, frontier quality meant frontier pricing. Opus 5 breaks that link. For a founder deciding whether an AI feature is even economically viable, halving the token bill while keeping near-flagship quality is the entire game. I dug into that economics problem in why most AI agent projects fail.
Third, it is the most aligned model Anthropic has shipped recently. In its own behavioral audit, Opus 5 scored 2.3 on overall misaligned behavior, the lowest of its recent models. If you are putting a model in front of customers, fewer off-the-rails moments is worth as much as a benchmark point.
One honest caveat, because you deserve it. Anthropic led its announcement with benchmarks like Frontier-Bench and ARC-AGI 3 rather than the usual SWE-bench Verified headline, and it published several comparison charts as images rather than tables. A number of third-party write-ups quote a specific SWE-bench Verified figure for Opus 5, but I could not confirm one on Anthropic's own page, so I am not going to repeat it as fact. Everything charted above traces back to Anthropic's announcement or to an independent evaluator.
How to get the best results from Opus 5
The model is only half of the outcome. Opus 5 behaves differently from Opus 4.8, and a prompt tuned for the old model leaves real quality on the table. The following is drawn from Anthropic's own guidance for the model, plus what I have seen building with it.
Thinking is on by default now. On Opus 4.8, leaving the thinking setting out meant no thinking. On Opus 5 the same request runs adaptive thinking automatically. That is good for quality, but your max_tokens cap now has to cover the thinking and the answer together, so a tight limit that worked before can truncate a response mid-thought. Give it more headroom.
Do not reach for the highest effort setting by reflex. Opus 5 exposes five effort levels, from low to max. Anthropic recommends starting at xhigh for coding and agentic work, and high for other demanding tasks, then sweeping down, because low and medium are unusually strong on this model and often beat the top settings of the previous generation at a fraction of the tokens. At xhigh or max, give it a large max_tokens, starting around 64,000, so it has room to think and act.
Give it the whole task up front. Opus 5 is at its best on long, autonomous runs where it receives the full specification in one clear brief and is then left to work, rather than being fed the task piecemeal across many turns. Front-load the goal, the constraints, and your definition of done.
Tell it to be concise, in words. Opus 5 writes longer responses and longer files than its predecessor by default. Turning the effort dial down does not reliably shorten the output, but an explicit instruction to keep responses focused does. In Anthropic's own testing, a short conciseness instruction cut response length by around 20%.
Delete your "double-check your work" instructions. This one is counterintuitive. Opus 5 verifies its own work without being asked, so the self-check and "add a verification step" instructions that helped older models now push it to over-verify and burn tokens. Removing them improves the result with no loss of quality.
If you want the full, plain-English version of how to prompt this model well, I wrote a separate guide: how to prompt Claude Opus 5.
Two smaller wins worth knowing. Prompt caching now starts at a 512-token minimum instead of 1,024, so short prompts you had written off as uncacheable will now cache with no code change. And if you need speed, fast mode runs Opus 5 at about 2.5 times the default output rate for twice the base price, which is a reasonable trade when latency is what your users feel.
Where this leaves you
The pattern that matters here is not "a new model dropped." It is that near-frontier capability just arrived at half the flagship price, with better alignment and a real step up in agentic work. That changes which AI features are worth building, because the ones that were too expensive to run last month may pencil out this month.
That is the calculation I help founders make: not "should we bolt a chatbot on," but which model, at which effort, wired into which part of the product, actually returns more than it costs. If you are weighing an AI feature and want a straight answer on whether it is worth it and how to build it well, see what I work on or book a call.