AI & Automation Briefing - October 28, 2026
OpenAI and Anthropic Have Cut LLM Prices by Up to 80% in Four Months
Between late June and late September 2026, OpenAI and Anthropic repeatedly cut model prices across their lineups. OpenAI dropped GPT-5.6 Luna by 80% on July 30 after it would have lost on cost-efficiency to DeepSeek V4 Pro at its original price point. Anthropic made its temporary Sonnet 5 launch discount permanent in August. On September 1, Anthropic released Fable 5.1 with a 75% cut to cache read costs, which translates to a 70-75% total cost reduction for agentic coding workloads where cache reads account for 96-99% of token spend. On September 22, OpenAI released GPT-6 Sol at $2/$10 per MTok input/output, half the price of the already-discounted GPT-5.6 Sol, and cut Luna pricing in half as well. The same day, Anthropic released Claude Opus 5.5 at $4/$20 per MTok, delivering a 40-60% total cost reduction versus Opus 5 with better benchmark performance. The pattern across both companies points to competitive pressure and improved inference efficiency driving prices down faster than most operators anticipated.
Research Finds Most Agent Memory Systems Underperform Simple RAG
A paper from researchers at the University of Illinois Urbana-Champaign introduces the concept of "memory tax" to explain why dedicated agent memory systems so often fail to outperform naive retrieval-augmented generation. The study ran large-scale evaluations across multiple benchmarks and model backbones, finding that many memory systems deliver negative net value compared to pulling directly from raw history. The researchers frame memory's impact as a tradeoff: memory can reorganize or distill information in ways that help a task, but those same transformations can also remove or obscure relevant evidence, producing the memory tax. The paper argues the field should shift focus from raw performance benchmarks to net value measurements when evaluating memory designs. Code and results will be open-sourced.
AI Agents Are Increasing Workload, Not Reducing It
A principal engineer turned VP of Technology at Bold.org tracked his GitHub contribution history from 2019 through October 2026 and found that AI agents drove his activity back to early-career levels, despite his role shifting away from hands-on coding. His contributions dropped steadily from 2,711 in 2020 to a low of 1,189 in 2023 as his responsibilities moved toward architecture and leadership. After adopting AI agents for engineering work in 2026, his count reached 2,102 by late October, surpassing every full year from 2021 to 2025. August 2026 became his single busiest month on record at 427 contributions, well above the previous high of 319. His median weekday contribution count also doubled, from 3 to 4 in prior years to 7 in 2026. He attributes the increase not to harder individual tasks but to managing multiple agents running in parallel, each requiring periodic context-switching to review plans, approve diffs, or interpret outputs. The pattern is relevant for operations leaders deploying AI tools: parallel agent workflows can expand total task volume and cognitive load even as they lower the barrier to starting individual tasks.