A Tale of Two Claudes

TL;DR: We pitted Anthropic's two flagship models against each other on a sliding scale of normie tasks—forms, spreadsheets, trip planning, research. Claude Sonnet 4.6 isn’t better than the premium Opus 4.6 model—but it is cheaper and more than good enough for most. Opus still wins when precision, judgment, or careful planning matters.

Some background: AI model updates aren't just about raw capabilities anymore—they’re about value for money. Case in point: Anthropic's Sonnet 4.6, released last Tuesday, claims to deliver near-Opus performance at a fraction of the cost. That matters, because the latest Opus costs more per token—the unit AI models use to measure and process text—so even users on Anthropic's Max plan burn through their limits quickly. (For every three Opus queries, you can do about five with Sonnet for the same price.) Here’s how they stacked up in our testing:

Clear Sonnet wins:

• Filling out a multipage auto insurance quote form. I told it to make up placeholder info to get an online quote from Progressive. Sonnet was a few minutes faster; Opus seemed to get tripped up by the birth date field.
• Finding current news on X. Both finished in roughly the same amount of time, but Sonnet surfaced actually recent news, while Opus brought up topics from as far back as January 2025.

Clear Opus wins:

• Vibe coding apps. I wanted to make a tool that autorenames my screenshots based on what’s in the image. Sonnet asked questions, but delivered an app that didn’t work. Opus caught bugs and gave me a working tool on the first try.
• Processing massive docs. Sonnet handled a single PDF without issue, but froze when asked to download all of the Mag 7's annual reports at once. Opus did it in about five minutes—and it actually grabbed PDFs; Sonnet got HTM files. (Once the files were in hand, both did a comparable job on the analysis itself.)
• Philosophy, ethics, and history questions. Opus took clearer positions and followed instructions more precisely. When asked what the first printed book was, Opus said the Diamond Sutra. Sonnet led with the Gutenberg Bible, then added caveats on “movable type” before landing on the correct answer.

Toss-ups (where you’re probably fine to use Sonnet):

• Spreadsheet analysis. Opus was faster because it clarified how much detail I wanted first; Sonnet asked no questions, but because it pulled a broader data set, analysis ended up being more detailed and interesting.
• Creating custom Skills. I asked them to make a Cowork Skill to log how long each prompt and response took within a task. Both got the job done; Opus's formatting was slightly neater.
• Group trip itinerary. Sonnet’s output involved more color and a peppering of cringe emojis. Opus avoided an incorrect assumption from a screenshot I provided. Both included roughly the same info and activity recs.

The verdict: Think of Sonnet as the Goldilocks option of the Claude family, sitting between Haiku—the cheapest model—and Opus. Opus is more meticulous, self-aware, and reliable, but that level of precision is expensive and, for most tasks, unnecessary. Only upgrade when: You can’t afford silent errors, you’re handling complex code or multidocument reasoning, or you want the model to challenge your assumptions.

What’s next: Frontier labs used to compete on pure intelligence. Now they’re competing on value per token. Sonnet doesn’t dethrone Opus—but it doesn’t need to. It’s “good enough” at a price that makes it the practical default. In that sense, the future of AI may not belong to the smartest model. It may belong to the one that delivers 90% of the performance at 60% of the cost. —WK