Breaking01 / AI
More dreaming lifts Claude's wiki quiz score from 13% to 48%
In company-published figures, combining search with the wiki scored 61%, and Anthropic did not disclose the experiment's total cost.

Anthropic said on 6 Nov 2026 that Claude's score on a wiki-based question test rose from 13% to 48% of graded criteria, a company-published figure, after it scaled up dreaming, a process in which the model reviews its records to organise its memory.
The research post describes a test on about 350,000 internal document files from Stainless, a company Anthropic acquired. Claude Opus 5 agents wrote and corrected wiki pages and merged duplicates across successive sessions, as dreaming grew from two sessions to 128. The score measured how many grading criteria Claude met across 25 questions.
Search combined. The score using both search and the wiki rose from 54% to 61% as the wiki grew, according to company-published figures. The search tool alone scored 53%. Using search and the wiki together also saved about 20,000 tokens per question.
Anthropic wrote that telling Claude what memory is for, without setting out how to use it, helped. When given a rule to record only failures, Opus 5 and Opus 4.7 lost much of the performance gain that memory gave them.
The post also compares memory use between models, using company-published figures from one Pokémon Red run each. At about 10,000 steps, Claude Sonnet 3.7 had written 413 files across 225 folders and earned one badge, a marker of progress in the game. Claude Fable 5 had written nine files in two folders and earned four badges.
Anthropic did not disclose the total cost of the 128-session dreaming run in its description of the Stainless experiment.