Sonnet 5.5 Tops Opus 5.5 on 1 of 8 Anthropic Benchmarks

A Moonshots podcast clip says Sonnet 5.5 jumped to 70% on Terminal-Bench and beats Opus 5.5 at 66.4% for half the price. Anthropic’s own pages list 70.6% for Sonnet 5.5 against 66.4% for Opus 5.5 on Terminal-Bench 4.0, and also show Opus 5.5 ahead on the other seven rows of the same table, with cache-read prices identical. We transcribed the clip ourselves with automatic speech recognition.

What the Clip Says

A speaker on the Moonshots podcast, in episode 299, says Sonnet 5.5 “jumped from 10% to 70%” on Terminal-Bench 4.0 “in a single generation,” and that it beats “Anthropic’s top model Opus 5.5 at 66.4% for half the price.” The speaker adds that it is the first Sonnet launched with cyber safeguards, a phrase taken from the clip’s published transcript, because this publication’s own recognition caught only the start of that sentence.

This publication transcribed a clip of the passage with automatic speech recognition, so the wording may contain errors. It checked each figure against Anthropic’s own pages for Sonnet 5.5 and Opus 5.5.

What Anthropic’s Table Says on Terminal-Bench

Anthropic’s Sonnet 5.5 page lists Terminal-Bench 4.0 at 70.6% for Sonnet 5.5, 10.3% for Sonnet 5 and 66.4% for Opus 5.5. The clip’s figures match. This publication’s own arithmetic: the gap between Sonnet 5.5 and Opus 5.5 is 4.2 points, and the rise from Sonnet 5 is 60.3 points.

Anthropic’s Opus 5.5 page says its Terminal-Bench 4.0 result is reported at xhigh effort, with a standard error of plus or minus 2.6 points for Opus 5.5 and 1.6 to 2 points for the other Claude models.

The Other Rows

The same table has seven more rows, and on all seven Opus 5.5 scores higher. FrontierCode 1.1 reads 46.2% at Max effort and 52.1% at Xhigh for Sonnet 5.5, against 54.4% for Opus 5.5. CursorBench 4.0 reads 55.5% against 57.8%, and Humanity’s Last Exam with tools 64.5% against 67.7%.

OSWorld 2.1 reads 80.1% against 81.8%, GDPval-AA 1844 against 1846, AA-Briefcase 1811 against 1822, and Chartography without tools 61.6% against 64.4%. This publication’s own arithmetic: Sonnet 5.5 is ahead on one of the eight rows and behind on seven.

Anthropic writes that Opus 5.5 “remains clearly stronger at complex, open-ended work requiring sustained judgment.” It also says that on CursorBench Sonnet 5.5’s best score is within about two points of Opus 5.5.

The Half-Price Claim

Anthropic prices Sonnet 5.5 at $2 per million input tokens, $10 per million output tokens and $0.20 per million tokens for cache reads. It prices Opus 5.5 at $4 and $20 per million input and output tokens and $0.20 per million for cache reads.

This publication’s own arithmetic: the input and output prices are half, and the cache-read price is the same. Anthropic frames cost per task, not per token, saying Sonnet 5.5 costs up to 30% less per task than Sonnet 5, and that it complements Opus 5.5 best at lower effort settings, where it costs less per task.

Cyber Safeguards and “Top Model”

Anthropic says that because Sonnet 5.5’s cybersecurity capabilities are comparable to Opus 5’s, it is “the first Sonnet model to launch with cyber safeguards.” It says higher-risk cybersecurity tasks will visibly fall back to Sonnet 5.

On “top model,” Anthropic’s news listing says Opus 5.5 performs at the level of Claude Fable 5.1 on most work, and its Opus 5.5 page compares it with Claude Mythos 5.1 on biology and cybersecurity. This publication draws no ranking from that.

The Structural Read

The Moonshots clip says Sonnet 5.5 “jumped from 10% to 70%” on Terminal-Bench 4.0 “in a single generation,” and that it beats Opus 5.5 at 66.4% for half the price. Anthropic’s Sonnet 5.5 page lists Terminal-Bench 4.0 at 70.6% for Sonnet 5.5, 10.3% for Sonnet 5 and 66.4% for Opus 5.5, so the clip’s figures match. Anthropic’s Opus 5.5 page says its Terminal-Bench 4.0 result is reported at xhigh effort, with a standard error of plus or minus 2.6 points for Opus 5.5.

The same table has seven more rows, and on all seven Opus 5.5 scores higher, for example 54.4% against 52.1% on FrontierCode 1.1 at Xhigh effort and 81.8% against 80.1% on OSWorld 2.1. Anthropic writes that Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment.

On price, Anthropic lists Sonnet 5.5 at $2 and $10 per million input and output tokens and Opus 5.5 at $4 and $20, with cache reads at $0.20 per million tokens for both. Anthropic frames cost per task rather than per token, and says Sonnet 5.5 costs up to 30% less per task than Sonnet 5.

Three Implications

THE COUNT OF ROWS This publication’s own arithmetic is that Sonnet 5.5 is ahead on one of the eight rows in Anthropic’s table and behind on seven.

THE PRICES The input and output prices are half and the cache-read price is the same, by this publication’s arithmetic. Anthropic frames cost per task rather than per token.

WHAT THIS PUBLICATION DID NOT CHECK This publication did not read the effort setting behind Sonnet 5.5’s 70.6%, did not retrieve the table’s footnote text on the Sonnet page, and did not check any independent leaderboard.

What Is Not Established

This publication did not read the effort setting behind Sonnet 5.5’s 70.6%, did not retrieve the table’s footnote text on the Sonnet page, and did not check any independent leaderboard. Every benchmark and price is Anthropic’s own.

The clip is accurate on the figures it quotes and silent on the other seven rows. Nothing here says which model to choose, and nothing in this piece is investment advice.

Business Engineer Framework

The Map of AI — The Stack Beneath the Applications

The Map of AI places more than 200 companies across nine layers of the stack, from silicon to application. It shows the layers that sit beneath the applications people use every day.