The long-document table now shows only max_len=8192 (assets/long_context_8192.png),
and the Quickstart recommends it: strong up to about 4,000 tokens of text,
more variable beyond. Code is unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
research/scripts/bench_long_context.py places 20 support requests in 8
languages after 0 to 7,000 tokens of unrelated text and scores them at the
default 1,024-token limit and at max_len=8192. The table, rendered to
assets/long_context_table.png and shown below Installation, shows the higher
limit is reliable up to about 4,000 tokens (17 to 18 of 20 against 7 of 20)
and unstable beyond it (8 to 17 of 20). The Quickstart callout now says so,
recommends max_len=4096, and corrects the earlier note that accuracy held all
the way to 8,192. Code is unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
