Reddit deep dive says token compression can raise OpenAI API bills, not cut them

KitchenAmoeba4438 · reddit · 2026-07-22

• A Reddit deep dive argues that token compression tools do not automatically save money on OpenAI APIs.
• With GPT-5.6, cache writes cost 125% of normal input while cache reads cost 10%, so modifying an already cached prefix can become far more expensive than simply reusing it.
• The author says a change must remove more than 92% of the invalidated suffix to pay for itself on the next request.
• For RTK, the post argues its “tokens saved” metric is misleading because it cannot see actual API billing details, cache state, or cache breakpoints; one benchmark reportedly showed 96.2M tokens saved while the bill rose 7.6%.
• Headroom is described as more capable in theory because it proxies requests, but the current implementation still appears unable to account for GPT-5.6’s explicit breakpoint fields or write premium correctly.
• The conclusion: measure cost per successful task, not token counts, and only certain compression setups can actually save money.

Original post →

More from Infra

• SmolVM open-sources persistent computer infrastructure for agents that outlive chat sessions — aniketmaurya · 2026-09-11
• PyTorch Day Korea 2026 launches first offline conf, CFP closes Sept 13 — PyTorch · 2026-09-11
• Local LLM server dilemma: 4x CMP-170HX (price up 53% in 20 days) vs Mac Studio M5 Ultra — rumboll · 2026-09-11
• llama.cpp lands Flash Attention tuning for RDNA4, big prefill gains on AMD — pmttyji · 2026-09-11
• Your p99 latency benchmark may be lying: a deep dive into coordinated omission — Franc0Fernand0 · 2026-09-11
• Running MiniMax H3 on 12GB VRAM: quantization, Turbo LoRAs and attention backends compared — Possible_Mood676 · 2026-09-11

Infra channel →