Samwise Tech/AI/Robotics Newsletter
Wednesday, September 23, 2026
OpenAI Releases GPT-6 Sol and Luna, Cutting API Costs by 50%
OpenAI released two new models, GPT-6 Sol and Luna, on September 22, expanding its GPT-6 generation with updated mid-tier options. Sol targets long-horizon coding and agentic work, priced at $2 per million input tokens and $10 per million output — cutting costs by 50% or more versus GPT-5.6 equivalents. Luna, designed for lighter tasks, costs $0.10 input and $0.50 output per million tokens. Sol achieves 88.8% on Terminal-Bench 4.0 in max mode. OpenAI says both models produce fewer errors than predecessors; the company describes Sol as built on the same underlying architecture as its frontier Astra system.
Sources: TechCrunch
Anthropic Releases Claude Opus 5.5 With Lower Prices and Fable-Level Performance
Anthropic released Claude Opus 5.5 on September 22, its first new model since CEO Dario Amodei publicly endorsed calls to slow frontier AI development to match alignment progress. The model achieves 66.4% on Terminal-Bench 4.0 and 54.4% on FrontierCode 1.1, outperforming the larger Fable 5.1 on most agentic benchmarks. Output tokens are priced at $20 per million, down from $25 for the previous model — a 60% overall API cost reduction versus Fable. TechCrunch reports Opus 5.5 succeeded in informal agentic tasks where Fable 5.1 repeatedly failed, despite using fewer active parameters.
Sources: TechCrunch
C2C Lets AI Models Communicate Through KV Caches, Cutting Handoff Latency
Researchers from Tsinghua University and collaborators published Cache-to-Cache, or C2C, a technique allowing AI models to share context directly via key-value caches rather than regenerating text for each handoff. Experiments show C2C improves accuracy by 3.1 to 5.4 percentage points over conventional text-based model handoffs and reduces latency across tested model pairs. The approach uses a trained cache fuser that maps one model’s KV-cache representation into another’s vector space — computationally cheaper than recomputing context. The paper appeared at ICLR 2026; researchers announced in September plans to release a full agent-managed serving implementation.
Sources: VentureBeat
Arm Launches Robotics Capability Framework to Reduce Industry Fragmentation
Arm unveiled its Robotics Capability Framework on September 23, a standardized industry program aimed at reducing fragmentation across the robotics software and hardware ecosystem. The framework establishes shared definitions for robot capabilities, allowing manufacturers, integrators, and software developers to communicate requirements without proprietary translation layers. Arm says the initiative covers humanoid robots, industrial arms, and autonomous vehicles equally, and is opening the framework to industry partners for review and adoption. Arm, whose processor designs power the majority of mobile devices worldwide, has increasingly positioned itself as core silicon infrastructure for the emerging physical AI deployment wave.
Sources: Robotics & Automation News
Distyl AI Raises $175M Series B at $1.8B Valuation for Enterprise Workflow Automation
Distyl AI announced a $175 million Series B on September 23, valuing the enterprise AI workflow startup at $1.8 billion. Khosla Ventures and Lightspeed Venture Partners led the round, bringing the company’s total funding to more than $200 million. Distyl builds AI-native workflow automation for large enterprises, generating and deploying agentic processes that integrate directly into existing ERP and CRM systems rather than overlaying legacy automation tools. The raise reflects strong investor appetite for enterprise AI companies targeting workplace process automation, one of the most active funding categories in 2026.
Sources: TechCrunch
Nexstrom Raises $12M Seed to Commercialize 2D Semiconductor Materials for Post-Silicon Chips
San Francisco startup Nexstrom emerged from stealth on September 22 with a $12 million seed round to commercialize two-dimensional semiconductor materials known as transition-metal dichalcogenides, or TMDs. Unlike silicon, which is approaching physical scaling limits, TMDs can be thinned to a single atomic layer while maintaining electrical properties, enabling denser, lower-power chips. Nexstrom’s flagship product, North Star, targets semiconductor foundries seeking post-silicon substrate materials for next-generation logic devices. Researchers have studied TMDs experimentally for over a decade; Nexstrom contends that recent advances in crystal growth yield rates now make commercial production viable.
Sources: TechCrunch
Discord Launches Mandatory AI Age Verification Despite Community Backlash
Discord launched mandatory AI-powered age verification on September 22 for access to age-restricted content channels, drawing significant community backlash over privacy concerns. The system uses machine learning to classify users based on account signals and behavioral data; Discord says more than 90% of users pass the check automatically without submitting government identification. Users the algorithm cannot classify must verify age through a third-party identity service. Critics objected to Discord collecting biometric or identity data and to the opacity of the ML classification criteria. Discord said the rollout responds to emerging online safety regulations targeting underage access to adult content on social platforms.
Sources: TechCrunch
⚡ Tech Pulse — September 23, 2026
Frontier (TB-4.0): GPT-6 Sol 88.8% · Gemini 3.7 Flash 85.8% · Claude Opus 5.5 66.4%
Open Source (SWE-Bench Pro): GLM-5.2/Z.ai 62.1% · MiniMax-M3 59.0% · Kimi K2 Thinking (new)
Small 15–50B: Qwen3.8-27B · Nemotron-Cascade 2 (30B MoE) · Bonsai 2-27B (PrismML)
Edge 0–15B: Qwen3.5-9B · LFM2.5-2.6B (Liquid AI) · Granite 4.0 Nano (IBM)
AI by Market Cap: NVIDIA $4.6T · Microsoft $3.8T · Alphabet $3.25T
Robotics by Value: Intuitive Surgical $195B · Figure $39B (pvt) · Unitree $36B (pvt)
Every newsletter preserved and searchable
Curated by JD · samwise.agency