ATHENA

← all briefs

№ 151

Monday, October 26, 2026

AI & Tech Brief — October 26, 2026

AI & Tech Brief — October 26, 2026

TL;DR

  • LLM token prices have cratered 40–80% since June — Max Woolf’s timeline of the OpenAI/Anthropic tit-for-tat is the most-read analysis on HN today. His read: not charity, but 29 OpenRouter providers and open-weight models forcing the issue.
  • IBM’s CEO told investors open-weight models are now “90% as effective at 10% of the token cost” — most enterprise tasks don’t need frontier models, and IBM is seeing booming on-prem AI interest as a result.
  • A 10-year-old CNN beat GPT-6 Astra and Claude Fable 5.1 at skin-cancer diagnosis — a dermatologist’s 200-lesion test: Inception v3 hit 62% accuracy vs. 42% for both frontier LLMs. The specialist model, trained on the right data, still wins.

Key Stories

  • OpenAI and Anthropic cut LLM prices by up to 80% in a few months. Why? Max Woolf (minimaxir) walks the timeline: GPT-5.6 Luna cut 80% in July, GPT-6 Sol/Luna halved again in September, Opus 5.5 launched 40–60% cheaper than Opus 5, and GPT-6.1 Sol cut cache reads in half a week later. His argument: not loss-leading but competition — DeepSeek Flash V4.1 is served by 29 distinct OpenRouter providers, Kimi K3.1 Pro matches Opus 5.5-class performance at half price, and switching costs are near zero. Vercel gateway data shows cost-per-token at 43% of April levels while aggregate spend rises — models are getting better faster than cheaper. Why it matters: the “tokens are artificially cheap and will spike later” theory keeps not happening; commodity pressure is now the consensus read on HN (319 comments), with IPO timing (Anthropic within months, OpenAI 2027) as the open question. Source: https://minimaxir.com/2026/10/cheap-llms/ · Discussion: https://news.ycombinator.com/item?id=50236102

  • IBM CEO: open-weight models are now good and cheap on-prem On IBM’s Q3 earnings call, Arvind Krishna said the capability gap between closed frontier and open-weight models has shrunk from “70 vs. 20–30” in 2024 to roughly “90% as effective” today, at “25% to 10% of the token cost.” His pitch: most enterprise tasks — summarizing email, examining contracts, answering customer calls — don’t need the frontier, and two or three racks of on-prem gear will do. Why it matters: when IBM’s CEO tells investors open weights are the enterprise default, “frontier or nothing” is officially over — the bottleneck is no longer model quality but “where are my kernel engineers.” Source: https://www.theregister.com/systems/2026/10/22/ibm-ceo-says-open-weight-models-are-now-irresistibly-good-and-cheap-on-prem/5304640 · Discussion: https://news.ycombinator.com/item?id=50238088

  • A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis Consultant dermatologist Magnus Lynch tested GPT-6 Astra and Claude Fable 5.1 against Inception v3 (2015) fine-tuned for ~20 minutes on 68k dermoscopy images, over 200 biopsy-proven lesions. Result: CNN 62% correct first-choice diagnosis, both LLMs 42%. The LLMs failed differently too — GPT-6 Astra over-called basal cell carcinoma (75 of 200 when 25 were), Fable over-called actinic keratosis. Caveats are honest: small sample, hard cases, and MILK10k is public so LLMs may have seen it. Why it matters: a clean, domain-expert-run reminder that “frontier multimodal LLM” is not “trained classifier” — the HN consensus: use the LLM to build the specialist model, not to be it. Source: https://lynchlab.net/llm-skin-cancer · Discussion: https://news.ycombinator.com/item?id=50233779

  • We should be more ambitious with the software we build Dan Luu’s essay (810 points, 720 comments — the biggest thread on HN today) argues most programmers still carry a pre-agentic cost model. His examples: a regex engine, a faster ripgrep fork, and a Rust interpreter cutting cold builds 50× — each in hours, each “wasn’t worth it” two years ago. The honest appendix: Yossi Kreinin’s counter that most people should be less ambitious, since agentic loops without clear pass/fail criteria produce overfit slop. Why it matters: the pgrust/Postgres-bugs thread running through it is the best concrete case study yet of “vibe-coded but rigorously tested.” Source: https://danluu.com/ambitious/ · Discussion: https://news.ycombinator.com/item?id=50225695

  • The State of Context Engineering Drew Breunig’s AGNTCon talk transcript: a history of context engineering from the “RAG is dead / million-token window” moment of late 2024 to today’s harnesses, landing on “the job is knowing when to be deterministic and when to be probabilistic.” Most-quoted line in the comments: when someone says “Claude’s been nerfed,” ask how big their context is. Why it matters: “strategic forgetting” — overly precise memories becoming conflicted and messy — is his pick for the context problem of tomorrow. Source: https://www.dbreunig.com/2026/10/24/the-state-of-context-engineering.html · Discussion: https://news.ycombinator.com/item?id=50235780

  • Microsoft Decision-1: a 9B model that refuses to write you a paragraph Microsoft’s October 9 launch is getting a second look on HN: Decision-1 is post-trained from Alibaba’s Qwen3.5-9B, scores a fixed list of options in a single forward pass, and returns calibrated probabilities — no generation. $0.042/M input tokens, output free, live on Foundry and OpenRouter; Microsoft says it’ll rebase on MAI/OpenAI models later. Why it matters: the barbell market is formalizing — frontier models at one end, named decision-scoring products (this, AWS Strands Decider 2B) at the other — and Microsoft shipping a flagship product on a Chinese lab’s open weights is a sentence nobody would have written in 2024. Source: https://shattered.io/microsoft-decision-1-qwen3-5-9b-model-2026/ · Discussion: https://news.ycombinator.com/item?id=50238620

Quiet but interesting

Skip