Cache Analyzer is described as a site that analyzes Claude Code session activity to compare Anthropic’s five-minute and one-hour prompt-cache TTLs, although the supplied snippets do not identify its creator or explain its implementation. Anthropic’s documentation says the one-hour TTL has a higher cache-write cost but can avoid reprocessing after longer idle gaps, while five minutes is generally cheaper for short, interactive bursts. The best choice therefore depends on observed context reuse and session cadence; Anthropic reportedly says Claude Code selects the TTL automatically rather than exposing a global user setting.
Scott’s “Prefix-Caching Economics” page already holds the substantive position that cache value depends on context reuse across agent runs, while “Agent Observability” covers deriving cost and latency decisions from session traces. Cache Analyzer is a directly relevant implementation example, but the supplied evidence establishes neither a new finding nor an actionable Claude Code control—indeed, TTL selection is reportedly automatic—so it does not yet change what Scott would build or argue.
ip:concept.prefix-caching-economicsip:concept.agent-observabilityradar:cache-hunter-prompt-cache-debuggingradar:wattage-claude-code-token-wasteradar:concept.prompt-caching
queries asked of Scott's wikis
- coding-agent session telemetry and cost optimization
- prompt-cache TTL economics for agent workloads
- adaptive caching for long-running coding agents
- agent harness observability and token-cost analysis
- context reuse across main agents and subagents
- latency versus inference cost in interactive coding
2026-09-05T07:26:09Z
Repeated refreshes now concern Portal’s delegation approach and article presentation, not Cache Analyzer’s TTL analysis. With no substantive analyzer evidence since discovery and no identifiable forthcoming validation, this case has faded rather than been disproved; unrelated optimization discussions should not keep it active.
2026-09-05T06:26:56Z
The refreshed Portal comments remain discussion of delegation and source indexing, not independent evidence for Cache Analyzer’s TTL comparison. No methodology, session traces, or actionable cache-retention choice has emerged; tangential optimization discussion should not sustain hourly review.
2026-09-05T05:25:11Z
The refreshed Portal discussion adds anecdotal objections to multi-model delegation and a source-indexing alternative, not evidence about cache retention. Cache Analyzer remains an unvalidated TTL-analysis example; discussion of separate token-saving techniques should not keep driving hourly reassessment.
2026-09-05T03:24:04Z
A newly quoted benchmark narrows Portal’s roughly 90% claim to Claude bulk-read tokens in four Java-monorepo scenarios, not demonstrated end-to-end cost savings or better cache retention. This adds specificity to a separate optimization technique but leaves Cache Analyzer’s TTL comparison unvalidated.
2026-09-05T02:23:36Z
The refreshed Portal discussion repeats the multi-model delegation explanation rather than adding cache-retention evidence or validating total-cost savings. Cache Analyzer’s TTL comparison remains unvalidated; tangential cost-optimization anecdotes do not warrant expanding or promoting this case.
2026-09-05T01:26:15Z
New comments suggest Portal’s claimed 90% Claude Code token reduction comes from delegating work to other models and token budgets, not improved cache retention; it therefore cannot be read as equivalent total-cost savings. This weakens its usefulness as supporting evidence for Cache Analyzer, whose TTL comparison remains unvalidated.
2026-09-05T00:29:16Z
Spotify’s 90% token-reduction headline suggests another potentially material context-handling implementation, but without the underlying mechanism or measurements it neither validates Cache Analyzer’s TTL comparison nor establishes an actionable Claude Code cache choice.
2026-09-05T00:22:56Z
evidence attached: hn.story.49571465 — Spotify's reported 90% Claude Code token reduction is independent evidence that request or context handling can materially change coding-agent cost.
2026-09-04T05:28:07Z
Refreshed comments only repeat that cache misses, token efficiency, and subscription quotas can offset cheaper cache reads. No session traces, controlled TTL comparison, or actionable Claude Code setting emerged, so the analyzer remains an unvalidated implementation example.
2026-09-04T02:28:47Z
Refreshed comments repeat the already-known distinction between cache-read pricing, cache-hit rates, token efficiency, and subscription quotas without supplying traces or a measured TTL comparison. The analyzer remains an unvalidated example rather than a developing finding.
2026-09-03T21:39:54Z
The new user reports suggest that cache misses, token efficiency, and subscription-quota accounting can overwhelm cheaper cache reads, strengthening the need for workload-level telemetry. They still do not validate Cache Analyzer’s TTL comparison or establish a user-actionable Claude Code cache choice.
2026-09-03T20:23:01Z
evidence attached: reddit.post.1w6hktb — User-reported costs and cache-miss behavior materially contextualize whether Claude Code cache settings deliver the claimed savings.
2026-09-03T05:25:20Z
The blended-cost model clarifies that cheaper cache reads may translate into much smaller total-bill savings once output tokens are included, but it still does not validate Cache Analyzer’s TTL methodology or produce a user-actionable Claude Code decision. The case remains a thin implementation example of economics Scott already tracks.
2026-09-03T05:21:47Z
evidence attached: reddit.post.1w5y52p — The blended-cost analysis materially contextualizes how cached-input pricing affects real coding-agent bills.
2026-09-02T07:36:19Z
The 36% savings headline reinforces prompt caching as a material cost lever but does not validate Cache Analyzer’s five-minute-versus-one-hour analysis, methodology, or applicability to Claude Code. The case remains a thin implementation example of economics Scott already tracks.
2026-09-02T07:22:01Z
evidence attached: hn.story.49532602 — This independent report of a 36% prompt-caching cost reduction corroborates that cache policy is a material coding-agent economics variable.
2026-09-01T02:27:43Z
The low-priority continuation anecdote is tangential to the analyzer and provides no methodology, measured cache savings, or evidence that users can control TTL selection. The case remains an unvalidated implementation example of already-known cache economics.
2026-09-01T02:22:42Z
evidence attached: reddit.post.1w3xktg — The reported low-priority continuation mode raises a concrete cache-retention and usage-cost interaction relevant to analyzing Claude Code cache economics.
2026-08-31T20:40:50Z
The forced recheck adds no evidence about methodology, measured savings, or user-controllable TTL behavior. This remains an unvalidated implementation example of already-known cache economics rather than a developing finding.
2026-08-31T20:33:37Z
grounded: known/low — Scott’s “Prefix-Caching Economics” page already holds the substantive position that cache value depends on context reuse across agent runs, while “Agent Observa
2026-08-31T20:30:28Z
case created — The released analyzer addresses a concrete prompt-caching economics decision not covered by an existing case.