The Opus 5 models feel bad by default and great once you know which levers to pull. Here is what changed and how to get on the happy path.
Don't have time to read?TL;DR
- The Opus 5 models on their default effort think more than you need, which makes them slow and eats tokens. Low effort is enough for daily driving.more ↓
- 5.5 is built for long unattended work in big codebases and ships with defaults to match.more ↓
- Opus 5 defaulted to high effort. 5.5 defaults to medium. At its default 5.5 thinks more per turn but finishes in fewer tokens and is faster.more ↓
- Thinking is always on and adaptive now. Delete any prompts that tell it how much to think and use effort instead.more ↓
- 5.5 has a bias for action, so tell it to look things up first.more ↓
- It runs loops longer now if you tell it when it should and shouldn't stop.more ↓
- You can give it a time budget to get complex work done quickly in parallel.more ↓
Opus 4.8 was easy. You could pick it up, leave it on high, and it was smart and fast enough to drive all day. Then Opus 5 came out and it felt terrible. It was smart. It was capable. It would get to a working answer. But it took forever to get there and it answered in the most roundabout way it could find.
It got bad enough that I stopped working and went looking for Anthropic’s own docs on how to prompt Fable, Opus 5, and Opus 4.8. Reading them, it all clicked. I finally understood why I was having the experience I was having, what problem these models were built to solve, and how they were meant to be driven. That is what I’m going to share here. By the end you’ll know what changed in the 5 class of models, what Opus 5.5 was built for, and the handful of settings and prompt habits that turn it from frustrating into something you’ll actually enjoy driving.
For context, my daily driver now is Fable 5.1 on low. I reach for Opus when I run out of Fable credits.
The Gemini problem
People never liked Gemini much, and I think it was for the same reason people don’t like Opus 5. Gemini was built to be steerable. It was very responsive to instructions, which made it capable, but it was weird if you weren’t steering it the way it expected to be steered. Left on its defaults it just felt off.
The Opus 5 family is the same. The default experience is not great, and you have to be more conscious about which levers you’re pulling. But there is a happy path, and it does not take much to find it.
What changed with the 5 class
The biggest change in the 5 class of models is thinking. These models were built to power agents that work on large problems in large codebases, using subagents to split the work up. To do that they had to get really smart and think a lot, and the Opus 4 models never had the amount of thinking that kind of work needs.
Anthropic also realized that not every problem is that complicated, so users needed a way to tune how much thinking a model does and what kind. That is the effort parameter. In the past you would reach for Haiku, Sonnet, Opus, or Fable to control how much thinking you got. Now you sit inside a single model like Opus and turn effort up and down instead.
Where I think they really screwed up was the default. Opus 5 shipped with effort on high. High is the right amount of thinking for working across a big codebase on a hard problem, with thorough validation and subagents. But most of what people hand these models while daily driving is not that complicated, and the result was a model that felt obtuse and sluggish. I think they chose high because, slow as it is, it almost always converges on a good answer. If you don’t know what you’re doing, you still get where you were trying to go. You just wait a long time for it.
The fix is to know what you’re trading when you turn the knob. On low it doesn’t think much and sometimes doesn’t think at all, but it is very fast. At medium and high it thinks more, validates more, and reaches for subagents more, somewhat regardless of how hard the problem is. At xhigh and max it reaches for a level of complexity I have yet to find a problem that needs.
The reality is that for daily driving I leave it on low, because even at the lowest amount of thinking these models are still very capable. And having it just shoot a response back when it doesn’t need to stop and think brings back some of the magic of the previous generations. You can have a conversation with it, and it will dip in and out of heavier thinking when it needs to. If something feels a little more complicated, especially something I think will need validation or would go faster in parallel, I bump it to medium. I almost never go to high unless I already tried medium and it didn’t work, which is about 1 percent of the time. In hundreds of hours of use I have yet to find a problem that needed xhigh or max.
The takeaway I hope you leave with is that we are moving away from a world where you switched between models a lot to one where you switch between effort levels a lot. Knowing that makes the Opus 5 family much more pleasant to work with.
What 5.5 is built for
Like its predecessors, 5.5 is built for long-running agentic work in big codebases. Anthropic’s examples are a 680,000-line migration finished in under a day and an engineering task that ran unattended for 18 hours. It was tuned for problems much harder than what most of us hand it on a normal afternoon, which is exactly why the defaults feel like overkill.
Effort on 5.5
Anthropic moved the default from high to medium, and they say 5.5 on medium matches Opus 5 on high. They even say that on some coding tests low comes close. They also say that at a given level 5.5 thinks more per turn than Opus 5 did, and A LOT MORE at xhigh and max.
Their recommendation is to start on medium. I think that is fine advice if you’re new. Starting on medium means you are probably getting more thinking than you need, so you can pay less attention to hallucinations and wrong answers. If you know what you’re doing, I would start on low and leave it there until something gives you a reason not to. Just know you will have to watch a little more closely, because about 2 percent of the time it is going to need help.
Other notable claims from Anthropic
- 30 percent faster than Opus 5, and uses fewer tokens
- Hallucinates less
- Writes more plainly
- Much better at reading charts, diagrams, and screenshots
- A bit cheaper, by about 20 percent per token
Driving tips
* A reminder that these are all early claims from Anthropic. Use them as a starting point, not as my field-testing results.
Thinking is always on now. You control it through effort. These models also use adaptive thinking, which means they pick how much to think based on the prompt. Effort sets the range, and on low that sometimes means the model just answers without thinking at all, which is great for daily driving. A byproduct of this is Anthropic’s recommendation to delete any prompts that tell the model how much to think or whether to think. That is a setting now.
The model’s running commentary is hidden by default. While it works, the model writes short notes about what it just found and what it’s doing next. Those notes used to come back as normal text. On 5.5 they are treated as thinking, and thinking is hidden unless you ask for it. If your tool only shows regular text, the model will look idle during a long task even though it is working the whole time. There is a display setting that brings the notes back.
It has a bias for action. The docs say this explicitly. If the model should look up information before it starts making changes, say so, because otherwise it will often just start doing things.
Loops are more steerable. If you want it to keep working, you can name the early stops you don’t want, like ending the turn on a summary that announces the next step instead of taking it. You can also name the stops you do want, like when nothing can move without your input.
You can give it a time budget. You can tell the model roughly how long a task should take, and it will pace itself to finish inside that. Why you would want this: subagents. For a long time, when a parent agent spun up subagents they would each run far longer than they needed to, which made parallel work take as long as doing it serially. With a time budget the model keeps its subagents short and still fans the work out. Anthropic says agent teams given a budget finish much sooner at the same quality. This is the most interesting change to me.
Things to watch
It runs more safety checks than Opus 5. There’s a new biology classifier and a new refusal for prompts that push it to print its internal reasoning. Everyday health questions and finding bugs in your own code are fine. I’ll be watching for overreach.
Where this leaves us
This model is very new, and everything above is Anthropic’s claims, not my measurements. I’m not trying to tell you how I will end up using it. What I hope to give you is a little insight into how I got these models to a place where I really enjoy working with them, and how I’d recommend starting your own exploration of Opus 5.5. Happy Clauding.
Sources
Find something fun? Hit me up on X.
How does it compare to Codex? No idea. I’ve used Sol and Astra a little and they seemed fine, not earth-shaking. I have a 20x Claude plan and a 1x GPT plan and haven’t found a reason to change that.