The model selector in Claude is one of those funny settings where I’ve come to learn most people fall into one of two buckets.
- Touch it once and forget.
- Turn on the most powerful model and start
The three models available to you right now (Haiku 4.5, Sonnet 4.6, and Opus 4.7) are meaningfully different tools. Not in the way that a hammer and a screwdriver are different, but in the way that a quick phone call, a proper working session, and a full discovery meeting are different. Each is the right choice in a specific context, and the wrong choice everywhere else.
Let's have a look at the models, what's changed with them, and of course, some practical examples. Oh yea, the hook!
YOU’RE LOSING $130,000 A MONTH ON TOKENS COST WITH THE WRONG MODEL!
What Each Model Actually Does
Haiku 4.5
Haiku 4.5 is built for speed. It's the fastest of the three, handles 200k tokens (roughly 150,000 words of context) in a single conversation, and keeps the cost low . If you're running something at volume, like:
- Summarizing a backlog of support tickets
- Classifying customer feedback
- Generating first-pass responses across a high number of records
Haiku is the right starting point. You'll notice it performs well in lighter thinking tasks and taking directions for retrieval tasks.
Sonnet 4.6
Sonnet 4.6 is where most professionals should spend the majority of their time. It balances intelligence and speed well, can handle 1M tokens (roughly 750,000 words of context). This is more than enough to:
- Work through an entire product spec
- Complete a full client history report
- Form a long research document
It also supports Adaptive Thinking (more on that below). For drafting, analysis, stakeholder communication, and most of the tasks that make up a busy day in B2B SaaS, Sonnet does the job without making you wait for it.
Opus 4.7
Opus 4.7 is the deepest thinker in the current family. It handles the same large context as Sonnet and brings more to complex, multi-step reasoning. Think of tasks like:
- Mapping a product decision across several competing priorities
- Synthesizing a large discovery document into a structured recommendation
- Working through a problem that requires holding a lot of moving parts in mind at once.
Opus is slower than Sonnet and carries a higher cost, so it earns its seat at the table on the tasks where accuracy and depth matter more than time. It also supports Adaptive Thinking.
What Has Changed Between Versions
This is really important to know, particularly how and why it affects your prompts. Especially if you are one of the people that fall into bucket 2 from above.
Opus 4.6 to 4.7, two behavioural shifts are worth knowing
- The newer version follows instructions more literally. If you ask it to review one section of a document, it'll review that section and not assume you wanted it to carry the instruction forward across the rest. That's good news when you're precise. But if your prompts have been written loosely, expecting the model to read between the lines, you'll want to tighten them up.
- The tone shift is also noticeable. Opus 4.7 is more direct than its predecessor and less focused on validating your framing before it gives you an answer. If you were used to warm, affirming responses signalling the model had understood the task, those signals are quieter now. The output is sharper; the warmth is gone.
When you hear noise about the performance of Opus 4.7, it's a lack of understanding on these changes. The reality is, if your prompting skills are not fundamentally strong, Opus 4.7 is not going to be the best model for you, and certainly not going to perform well in day-to-day tasks with ambiguity.
But of course your prompting skills are strong, because you're here, taking it brick-by-brick, which means you can choose Opus :)
Opus 4.7 is the unofficial model choice of Ralph Wigum.
Sonnet 4.5 to 4.6, less dramatic changes
- The context window expanded significantly, so you can now feed it much longer documents in a single conversation.
- The bigger shift is going from Extended Thinking to Adaptive Thinking
Haiku 4.5, not worth comparing to it's predecessor
I won’t cover the changes from Haiku since it updates less frequently than the Sonnet and Opus.
Just note that Haiku is on the 4.5 series, so it doesn't not have Adaptive Thinking, it has Extended Thinking. It's training data cutoff is also slightly older at July, 2025 vs January, 2026 for Sonnet and Opus.
What Is Adaptive Thinking?
Let’s first address the difference between Extended Thinking and Adaptive Thinking.
- Extended thinking is a feature where Claude works through a problem before responding, rather than answering immediately. In older versions, people would include telling Claude how much to think about something in the prompt.
- Adaptive Thinking is a feature where Claude makes that call itself, based on how complex the task actually is. For a simple question, it skips the extra step and responds quickly. For something harder, it takes its time. You don't have to manage the dial though.
In practical terms: if you notice Sonnet or Opus taking slightly longer on a complicated prompt than a simple one, that's often Adaptive Thinking in action. It's not a lag. It's the model reading what you've asked and adjusting how it approaches the answer accordingly.
💡Free Tip: If you spend a little bit of time reading the reasoning the model produces, you're going to start learning at an increased rate. This was a bit turning point for me personally.
Let's See It In Action
Here are 3 prompts to try with Adaptive on. The big learning point will be realizing how Claude is gauging the complexity of the task. If the reasoning is shallow, it’s because the prompt is asking something simple of the model. As always, try while on Incognito mode to make sure you’re not bringing in memory or instructions that would impact the outcome.
Prompt 1
What does MRR stand for?
Prompt 2
A customer just told me they're not renewing because they don't have the budget. But their product usage has gone up 40% over the last quarter. How do I respond to them?
Prompt 3
We're a 40-person B2B SaaS company with three CSMs managing around 180 accounts. We're thinking about introducing a tiered onboarding model where high-value accounts get a dedicated CSM and lower-value accounts move through a self-serve flow. What are the risks we should think through before making that change, and what would a sensible rollout sequence look like?
Example Outputs of Reasoning from Adaptive Thinking in Claude
A Simple Rule To Start With
There is so much more to unpack on this topic, but for now, the first brick is to pick your model the way you'd pick who to bring into a meeting. For a quick alignment check, you don't need your most senior person in the room. For a complex strategy session, Haiku isn't going to cut it.
Haiku for speed and volume. Sonnet for most of the day. Opus for the heavy thinking.