Anthropic describes Claude Opus 5 as a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at about half the price. Read that carefully. Close to. They could almost certainly have made it match Fable, and they did not. What you get is Fable-shaped capability at an Opus price, which is a good trade, but not the same as the top model for less.
The way this lands in real work is simple. A good habit is to start something new on the strongest model you have, get the process working, then move it down to Opus once the shape is settled. Opus 5 makes that handoff cost less quality than it used to.
Against GPT 5.6 Sol, Opus 5 takes novel problem solving by a wide margin, lands about even on agentic search at 90.8 percent, and falls a little short on agentic coding. That last one is the part most launch coverage quietly drops.
The number worth remembering is ARC-AGI-3, an evaluation built around problems the model has not seen before. Opus 5 scores roughly three times as high as the next best model. Opus 4.8 was not far off GPT 5.6 there; Opus 5 is clearly above both. Benchmarks are still benchmarks, and the gaps are often close enough that the chart flatters whoever published it, but a 3x margin on unfamiliar problems is not noise.
Ask the model itself and you get a short, unglamorous list.
| What changed | Why it matters |
|---|---|
| Swap tools mid conversation | On 4.8, changing the tool list rebuilt the whole prefix, so tool sets were effectively frozen for a session. Now they are not. |
| Automatic fallback on a refusal | A declined request reruns on a substitute model, routed by the kind of refusal, instead of you noticing and rerouting by hand. |
| 512 token cache minimum | Down from 1024. Short system prompts that never qualified for caching now do, with no code change on your side. |
| Thinking on by default | It verifies its own work unprompted, so you stop writing "and then check it" and start deleting the scaffolding that forced it. |
| Sub agent teams hold together | Agents stop overwriting each other, which is what finally makes writer and verifier patterns behave. |
One practical warning. This is the model you point at a pile of agents at once, and that is exactly how you burn a month of credits in an afternoon. On a lower tier plan, with something due today, do not turn it loose on a wide job. The other catch: tighter safety grades mean it can decline a request outright, which is the gap the automatic fallback exists to cover.
Anthropic's automated behavior audit calls Opus 5 their most aligned model to date, with the lowest rates of reckless or deceptive behavior and the strongest adherence to their published constitution. Worth saying plainly: nothing here suggests the model is misbehaving.
But it is an automated audit of a model by its own makers, and a more capable model is by definition better at producing whatever an evaluation rewards. That does not make the result wrong. It makes it the kind of result you hold loosely. The practical version of that caution is boring: be a little more deliberate about what you let it do unattended on your actual computer.
The security grades should lead and never do. Opus 5 is stronger than 4.8 on cybersecurity tests but still well behind Mythos 5 at developing exploits, which appears to be deliberate. It scores 79.4 on vulnerability identification, roughly on par with Mythos 5, while its exploitation success rate sits at 4. Low in absolute terms, and still four times higher than Opus 4.8. That reads less like a limitation and more like a company keeping a capable model inside a line it does not want to cross.
Most of what changed here is not something you feel as a smarter answer in a chat box. It is caching, fallback, and self verification: the plumbing that decides whether an AI process keeps working when you are not watching it. That is the difference between a clever demo and something that quietly does a job for you every week. If that idea is new, start with why AI needs a second brain to be useful.
The verdict is the least exciting sentence here and the most useful one. Smarter, same price. You do not need to change anything to benefit from it.
The free quiz takes 2 minutes and gives you a starting list for your business. No call, no card.
Take the free 2-minute quiz