Claude Opus 5.5 costs less to use: Anthropic reports a 40% reduction

Anthropic estimates as of Oct 1 that Claude Opus 5.5 costs 40% less than Opus 5 on typical tasks.

Opus 5.5 has almost stopped using em dashes, but retains thousands of writing patterns that appear more often in model output than in human writing.

Anthropic calls the new model a leader in agentic coding and intellectual tasks. The claimed 40% refers to the cost of completing typical tasks compared with Opus 5. There is a separate comparison for writing: on Oct 1, Graphite compared human and model-written articles on 9,974 shared topics. The human articles were published before ChatGPT appeared.

Em dashes have almost disappeared. Opus 5 used em dashes 2.92 times per 1,000 words. With Opus 5.5, the frequency fell to 0.015, a 99% drop. Word distribution also moved closer to human writing: Jensen–Shannon divergence fell from 0.064 to 0.052, a 19% reduction.

The patterns remain. Graphite counted 2,548 markers of model-written text in Opus 5.5, compared with 2,666 in Opus 5. That is a 4% reduction. Researchers counted words, phrases and constructions that appeared at least twice as often as in human writing.

Opus 5.5 used the word "dependable" 23 times more often than human authors, adjusted for text length. The phrase "this matters" appeared 116 times more often. In generated articles, the disappearance of em dashes was much more noticeable than the reduction in other patterns.

In Pangram's Sep 23 comparison, both models answered the same 1,000 prompts. Opus 5.5 used "in short" 100 times, while Opus 5 used it only 9 times. Average response length barely changed: 640.5 and 643.4 words, respectively.

In Annamika Aerts's test, the Pangram 4 detector classified 1,000 Opus 5.5 responses as follows:

- 997 were labeled entirely AI-generated. - 2 were labeled mixed. - 1 was labeled human-written.

Graphite shared tables of model-writing markers and the original articles on Google Drive via a link in the study.

Claude Opus 5.5 costs less to use: Anthropic reports a 40% reduction

Anthropic estimates as of Oct 1 that Claude Opus 5.5 costs 40% less than Opus 5 on typical tasks.

Opus 5.5 has almost stopped using em dashes, but retains thousands of writing patterns that appear more often in model output than in human writing.

In Annamika Aerts's test, the Pangram 4 detector classified 1,000 Opus 5.5 responses as follows:

- 997 were labeled entirely AI-generated. - 2 were labeled mixed. - 1 was labeled human-written.

Graphite shared tables of model-writing markers and the original articles on Google Drive via a link in the study.