After ~six months of integrating AI into our workflows at Lacerta Bio (plus building a few custom tools for ourselves and our clients), I’ve developed what I call the “intern-to-partner spectrum” for understanding Anthropic’s Claude models.
Each has earned its place in our toolkit, much like having different team members for different occasions.
Haiku: The Overenthusiastic Summer Intern
Haiku reminds me of that brilliant summer intern who comes in on Monday morning with seventeen ideas before finishing their double chai latte with oat milk.
Haiku is fast, eager, and very capable. It delivers with the enthusiasm of a young Golden Retriever chasing frisbees.
It’s perfect for rapid-fire tasks like summarizing piles of press releases or generating quick competitive intelligence briefs.
Want to review and critique a licensing presentation or a business plan? Want a summary from Investor Day? Haiku is surprisingly good at these kinds of tasks (better than in intern, to be fair).
The output quality? Solid B+ work that gets the job done without breaking the budget. You won’t get Pessoa-level prose (Haiku would not be able to write this post, for example), but it’s readable, functional, and worth the expense for the time saved.
Cost-wise, Haiku is your entry-level hire. It’s economical enough to throw at high-volume tasks without wincing at the monthly AI bill. And it won’t complain about having to work brutal hours.
Sonnet: The Reliable Mid-Level Associate
Sonnet occupies that sweet spot we all recognize. It’s the dependable senior associate who has enough grey hairs to know and ask the questions that matter.
Sonnet balances speed with sophistication. It’s good for more complex tasks like analyzing and comparing licensing agreements or preparing detailed competitive landscapes for board presentations.
Think of Sonnet as your go-to for substantive work that requires both analytical depth and practical business sense.
It won’t dazzle you with philosophical insights about the historical differences in drug pricing between two European countries, but it will deliver consistently strong analysis that you can confidently present to senior management.
Recent iterations of Sonnet now include Perplexity-like links to external sources of information. So far, each link has been spot on.
For novel tool development and coding, Sonnet is the go-to model for writing specification documents and suggesting prompts for Projects. Simpler coding exercises will run adequately on Sonnet. But for complex coding, especially when agents are involved, you need Opus.
Opus: The Seasoned Partner with Rapier-Like Wit
Opus is that colleague three doors down.
You know the one.
The colleague with three decades of experience who speaks in perfectly formed paragraphs and somehow makes complex regulatory strategy sound elegantly simple.
He can opine elegantly about both the differences between SSRIs and the use of cedar versus Sitka spruce soundboards.
When we’re structuring a tricky milestone payment scheme or need a deeper strategic analysis of a potential licensing opportunity, Opus delivers with the sophistication you’d expect from a senior partner.
Opus has an uncanny ability to ask the questions that might be slowly forming in your mind, but with a surprising level of clarity and precision.
And, like that colleague, once you get to know him, he will drop some amazingly funny quips. For example, A new entrant with a marginally differentiated GLP-1 agonist is bringing a knife to a gunfight.
Solid.
Writing code?
Bite the bullet and pay Opus to do it in Project mode.
With the right specifications, prompting, training data, and step-wise approach, Opus will deliver software that is at least functional, and usually at the first attempt. Opus is a must when writing coding that will involve agents.
The tradeoff? Like that seasoned partner, Opus commands premium rates. Flipping to an Opus chat is something you can do when the heavy lifting is needed. But like most things in life, it can be worth the expense.
Applications
For BD professionals, the challenge isn’t choosing one model versus another. It’s deploying the right tool for the right job.
These are different screwdrivers, or hammers, or saws, or whichever analogy you prefer. They are not larger or smaller versions of the same tool. They are different tools with different skills, different input needs, and different utilities.
Based on my limited experience to date, I generally recommend Haiku for volume work, Sonnet for daily analytical heavy lifting, and reserve Opus for the coding and decisions that truly matter.
But things are moving so rapidly in the AI/LLM space at the moment, that my analyses might be completely different next month!