Today's News
Sam Altman via AINews
OpenAI announced major price reductions across its GPT-5.6 model family: Luna dropped 80% to $0.20 per million input tokens and $1.20 per million output tokens (matching GPT-5.4's full intelligence at one-thirteenth the cost), while Terra fell 20% to $2/$12, and Sol gained a new Fast mode offering 2.5x speed at 2x price for equivalent intelligence. The dramatic efficiency gains stem from systematic improvements including 20% cost reduction via autonomous kernel rewriting, 15% efficiency gains from improved speculative decoding, and agentic harness improvements like prompt caching. This represents an annualized cost deflation rate of roughly 2000x for constant-level intelligence, accelerating sharply from the previously observed 1000x over 18 months—a concrete signal of the economics of frontier model scaling and inference optimization.
Amazon raises 2026 CapEx guidance to $220B and targets $1 trillion long-term AWS opportunity, with Jassy detailing AI infrastructure ROI math
Clouded Judgement by Jamin Ball
Andy Jassy disclosed Amazon's expanded capital intensity and market ambition: 2026 CapEx guidance increased to $220B, while AWS's long-term addressable market target expanded to $1 trillion (up from prior "few hundred billion" estimate). Jassy detailed the ROI mechanics: equipment achieves 3-year payback on 5-6 year useful life with 5-year customer contracts, while data centers have 30+ year lives that will cycle through 5-6 equipment generations, with each generation after the first printing pure profits. Critically, Jassy stated that even at $220B capex, Amazon won't have enough AI capacity to meet 2026 demand, with 2027 and 2028 reservations already booked. This directly addresses investor concerns about hyperscaler capex sustainability and signals confidence in sustained AI infrastructure demand through 2028.
Not Boring
Amazon's Zoox received NHTSA approval for limited commercial deployment of its novel steering-wheel-free robotaxis—the first purpose-built autonomous vehicle (not retrofit) approved for US commercial operation. The exemption permits Zoox to deploy up to 2,500 vehicles annually for two years and charge for rides, with service launching first in Las Vegas and expanding to additional markets. The vehicle features inward-facing seats, a 75 mph top speed, and exceeds equivalent federal motor vehicle safety standards, though NHTSA placed additional reporting requirements on crashes and operational issues, with authority to revoke the exemption if major safety problems emerge. This marks a regulatory milestone signaling the transition from retrofitting human-centric cars to designing vehicles purpose-built for autonomous operation.
Apple earnings via Techmeme
Apple reported strong Q3 results with $109.4B revenue, 27% net income growth, and product margins expanding to 40.1% despite the MacBook Neo launch. Services revenue now exceeds $30B, while iPhone grew 22% YoY and Mac grew 29%, suggesting the lower-cost MacBook Neo isn't cannibalizing margins as some feared. This signals Apple's AI hardware strategy is working and positions the company for a potential consumer upgrade cycle driven by personalized AI capabilities.
Amazon via Techmeme
Amazon reported AWS Q2 revenue growth of 37% YoY to $42.2B with operating income up 64%, while the company's custom silicon business reached a $25B+ annual chip revenue run rate. This demonstrates Amazon's vertical integration strategy is paying off and directly competes with Nvidia's dominance in AI infrastructure.
Google DeepMind releases Gemini Robotics 2 with unified whole-body control across diverse robot embodiments
Not Boring
Google DeepMind released Gemini Robotics 2, a new model family (VLA, embodied reasoning, and on-device variants) that controls humanoid robots like Apptronik's Apollo through full-body movement, dexterous manipulation, and multi-robot coordination using a single model checkpoint. This marks a meaningful shift toward unified perception-planning-control across diverse robot bodies, though movement speed and multi-finger dexterity still need improvement.
Cloud agents now account for 56% of merged PRs at Cursor, signaling material adoption in production software engineering
AINews
Cloud agents now account for 56% of merged PRs at Cursor (up from 10% in December), signaling material adoption of persistent agent environments in production software engineering workflows. Cognition added native GitHub stacked PR support and demonstrated Devin agents running macOS with Xcode simulator access, while Perplexity launched Projects with persistent memory infrastructure, though research evidence on memory's causal contribution to capability remains mixed.