Anthropic's context management post reports three figures, and the conditions matter more than the numbers:
- Memory tool plus context editing: 39% improvement over baseline
- Context editing alone: 29% improvement
- 84% reduction in token consumption in a 100-turn web search evaluation, where the baseline agent failed outright from context exhaustion
The stated condition is "an internal evaluation set for agentic search" on complex multi-step tasks. That set is not published, so none of these is independently reproducible — which does not make them wrong, and does mean they are evidence about that eval, not about your workload.
Two details worth knowing that get lost when these numbers are quoted. First, the 84% figure is partly an artifact of the comparison: the baseline failed, so this is closer to "the task became possible" than to a like-for-like efficiency gain. Second, the documentation itself carries no quantitative claims at all — only mechanism and defaults (clearing triggers at 100,000 input tokens, three tool uses retained). When the marketing post has numbers and the docs do not, the docs are usually the more careful artifact.