Good morning. The pricing story is turning: OpenAI and Anthropic have quietly cut inference prices by up to 80% since June, IBM’s CEO is telling investors open-weight models now do the job at 10% of the cost on-prem, and Apollo — again — says the broader economy isn’t seeing a dime of financial return. Elsewhere, a decade-old CNN is embarrassing frontier LLMs at skin cancer diagnosis, and a Fields medalist at Berkeley has given up grading homework.
Why OpenAI and Anthropic slashed prices 80% in a quarter. A post from minimaxir walks through the June–September price collapse at both labs, attributing it to Chinese competition (DeepSeek’s Flash V4 reportedly earns $38/GPU/hour at $0.60 per million output tokens), better inference efficiency, and the simple fact that GPU capacity is perishable like airline seats. HN commenters were split between “race to the bottom, someone goes bankrupt” and “this is exactly why both labs are suddenly so interested in regulation.” One pointed out cache-read reductions matter disproportionately for agentic coding, where they can be 70-75% of total cost.
IBM says open weights now hit 90% of frontier at 10% of cost. Arvind Krishna told investors that open-weight models have closed the gap enough to make 2-3 rack on-prem deployments compelling for routine tasks like email summarization and contract review, per The Register. Multiple HN commenters confirmed from personal rigs that they haven’t paid for a frontier model in months — one said GLM 5.3 with DeepSeek V4.1 as advisor feels Opus 4.6-level. Others noted Krishna is obviously talking his book, since IBM would like to sell you the Spyre accelerators to run all this.
Apollo doubles down: no AI payoff outside tech. Torsten Slok’s analysis of the S&P 500 minus the top seven tech names shows no measurable improvement in margins or revenue per worker from AI adoption, per TechCrunch and Business Insider. This echoes the finding we flagged yesterday, but the HN discussion added a sharper framing: companies buying expensive tokens expect profit margin expansion by osmosis, when all AI really does is let workers do more of the same mundane work faster. One commenter put it plainly — “AI is the anti-Midas golden finger.”
A 2015 CNN beats GPT-6 and Claude Fable at skin cancer. A fine-tuned Inception v3, trained in 20 minutes on a single H100, hit 62% accuracy across 200 biopsy-confirmed lesions versus 42% for GPT-6 Astra and Claude Fable 5.1 — crucially catching melanomas the LLMs waved through as benign. The HN consensus was “of course,” with the calculator-beats-LLM-at-arithmetic analogy getting trotted out several times. The genuinely interesting counter-question, raised by a few commenters: would the LLMs win if given patient history and allowed to ask follow-up questions? SkinGPT-4’s 80% dermatologist-rated accuracy suggests maybe.
Google’s AI Overviews are breaking publishers. The Verge’s column argues the symbiosis between Google and the open web — publishers optimize for SEO, Google sends traffic — is collapsing now that AI summaries answer queries directly on the SERP. Years of SEO strategy are being written off, and the traffic-dependent revenue model with it. No HN thread on this one, but the trend has been visible in publisher traffic reports all year.
Fields medalist stops grading math homework. UC Berkeley’s Richard Borcherds told the Daily Cal he now grades only on completion because AI has made correctness-grading pointless, and said OpenAI’s claimed Navier-Stokes solution has left him unsure whether to recommend math as a career at all. HN commenters pushed back toward in-class and oral exams as the obvious response, but a few noted a subtler loss: even students who copied homework used to engage with the material a little. AI removes even that minimum.
Context engineering’s second act. Drew Breunig’s State of Context Engineering argues that longer context windows haven’t solved the problem — they introduced new failure modes, and strategic forgetting is now the open question. The HN thread was unusually substantive: debates over cache invalidation when pruning turns, whether models should be given direct edit access to their own context, and multiple users reporting quality degradation past 200-300k tokens even on current models. One commenter liked the framing that engineers’ real job is choosing when to be deterministic (code) versus probabilistic (model).
Microsoft ships a Qwen-based decision model. Microsoft’s new Decision-1 is a 9B model built on Alibaba’s Qwen3.5-9B, designed only to score predefined answer options with calibrated probabilities — classification, routing, agent verification. Microsoft claims 35x speedups over general-purpose models for these narrow tasks. HN reception was tepid; the article reads like marketing copy and doesn’t mention Perplexity’s pplx-decider-v1.2 or other established competitors.
“Attention Is Not What You Need.” An arXiv paper proposes replacing self-attention with Causal Grassmann layers that encode token relationships as geometric subspaces, achieving linear scaling. Results land 10-15% behind transformers on Wikitext-2. HN tore it apart: one commenter flagged what looks like a data leakage bug (the pairing uses future tokens to predict future tokens), others dismissed the 13-18M parameter scale as meaningless, and several suggested the writeup itself was LLM-generated. A reminder that arXiv is not peer review.
That’s the morning. The through-line, if there is one: inference is getting cheap fast, open weights are catching up, and nobody outside the tech sector has figured out how to turn either into money.