The thesis is: Long-running coding agents make cache discipline and task completion efficiency primary drivers of cost.
The source reports that context per request in Claude Code grew 2.6x between March and September 2026, while Claude worked 3.3x longer on each prompt and interruptions fell 68%. Anthropic also says it cut input and output token prices by 20%, cached-read prices by 60%, and estimates Opus 5.5 costs about 40% less than Opus 5 for typical token-billed workloads. These are vendor-reported aggregate figures, not an independent benchmark.
The mechanism is: As agents repeatedly read large working contexts, cheaper cache reads, fewer cache misses, and cache inheritance by forked subagents reduce the marginal cost of sustained work. Why now is: Anthropic’s reported usage data says coding sessions are becoming longer and more context-heavy, increasing the share of spend exposed to those mechanisms. The first implication is: Engineering teams should measure cache-read share and task-level completion cost, not evaluate models solely by per-token list prices. The evidence boundary is: The source is Anthropic’s own product post; its cost estimates and performance claims may not generalize across codebases, workloads, contracts, or competing models.
The first watch condition is: Watch for independent task-level cost studies that compare cache hit rates, completion quality, and total spend across long-running coding workflows. If those measurements substantiate the vendor’s claims, context architecture—stable sessions, selective compaction, and reusable parent state—becomes a procurement and platform concern rather than a local prompt-engineering detail.