Claude Opus 5.5 on Hacker News: How to Read the Intelligence, Performance, and Price Columns

A Single Model Page, 274 Points

Artificial Analysis published a page on Claude Opus 5.5, with a title that puts three things side by side: intelligence, performance, and price. On Hacker News the link collected 274 points and 83 comments. In a stretch when leaderboards multiply faster than anyone can read them, a single-model page still reaching the front page says something about what people are actually trying to decide.

Intelligence, Performance, and Price Are Different Questions

Intelligence scores are usually a weighted result over a set of standardized tests. They describe a ceiling on controlled problems. They do not describe whether a model can land a real change in your repository. The numbers are useful because they are comparable, and limited for the same reason: test sets get optimized against.

Performance usually means time to first token and output speed. For batch jobs those two numbers barely register. For anything interactive they decide whether a user stays. Same output, two extra seconds before the first token, and it is a different product.

Price is the column people misread most. A per-million-token sticker price is a starting point. What lands on the invoice is the input-to-output ratio, cache hit rate, retries, failed calls, and how many rounds it takes to get an acceptable answer. A model that costs twice as much per token but gets there in one pass is frequently the cheaper option.

Stack the three together and the conclusion is rarely “one column wins.” It is “here is the column you can afford to give up.”

What the Thread Argues About

Eighty-three comments is not a large thread, but these discussions tend to cluster around the same disagreements: whether the benchmark tasks represent real workloads; whether a small version bump justifies migration cost; how far published pricing drifts from the actual bill.

For small teams, migration cost is concrete. Prompts need rewriting, evals need rerunning, edge cases need to be rediscovered. A .5 in the model name does not by itself justify pausing the work in front of you. For larger organizations the deciding factors usually sit elsewhere — compliance, regional availability, contract terms — and those carry more weight than a few leaderboard points once procurement gets involved.

There is a reason third-party pages get passed around instead of vendor announcements, and it is not that vendors lie. It is that a vendor’s own comparison table is built around the workloads where its model looks best, and readers know it. A neutral page has to make choices too — which benchmarks, which weights, which price tier — and those choices are exactly what a comment thread interrogates. Publish the methodology and the argument can at least be about the methodology.

There is also a read-only audience. They are not switching models. They want confirmation that their current choice has not fallen too far behind. That demand keeps pages like this one in circulation.

Which Column to Read

The value of a page like this is compression: given a budget and a latency requirement, which model clears the bar. It does not tell you how to write your project, and it does not decide whether a migration pays off.

The three columns map to three different questions. For the capability ceiling, read intelligence. For whether users will wait, read performance. For what the invoice looks like, read price — then recompute the first two against your own traffic. The 274 points suggest a lot of people are working through the same choice, and their spreadsheets do not match.