Anthropic's own cost guide includes a lever that cost 74% more
Brinvik · reddit · 2026-09-10
A lever in Anthropic's newly published cost guide — context editing and compaction — saved nothing and cost 74% more on a 20-issue run, yet saved 39% and 32% on longer runs. The sign flips with run length, meaning the guide is a set of bets on your workload rather than guaranteed wins, and someone must measure before changing anything.
Other figures from Anthropic's own benchmarks:
- Prompts written for Opus 4.8 cost 36% more per support ticket on Opus 5 with no accuracy gain; removing "verify twice" alone cut cost per ticket by a third on Opus 5.
- Agent loops read a median 84% of input from cache over a day of real traffic; top 10% of setups hit 94%+.
- Caching cut agent-loop cost by 2.7x to 5.3x.
- Without pauses, the five-minute cache default cost 15% less than the one-hour setting on Sonnet 5 and 11% less on Opus 5 — longer caches were pricier.
The author flags mismatched baselines (36% is Opus 4.8 vs Opus 5 untouched prompts; 14% is audited vs unaudited on Opus 5) and notes accuracy-restoring fixes (a retired thinking setting, contradictory rules, a hand-rolled scratchpad) each gave back 7-11 points — bigger than most cost levers.
More from coding & agent
- He benchmarks new models by having them build their own copilot — $28 of compute — bucolucas · 2026-10-27
- Founder generates 46 app promo videos in 4 days with Claude, skipping $11,500 in UGC fees — PrajwalTomar_ · 2026-10-27
- Should agents be allowed to change plans mid-task? Reddit debates the drift boundary — wistfulcracker84 · 2026-10-27
- HeyGen + Canva MCP turns slide speaker notes into an avatar presentation video — HeyGen · 2026-10-27
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27
- Cloudflare Computer adds agent interruption recovery and approval pauses in v0.7.0 — irvinebroque · 2026-10-27