Good morning. It's Monday, August 24, and we're covering open models closing the gap, calls for stronger frontier safeguards, and Anthropic’s push toward a potentially record-setting IPO.

Plus: a poll on AI devices and a Prompt of the Week for your next big plan.

Anthropic is weighing an IPO that could match or exceed SpaceX’s record $86.2 billion offering, signaling how quickly AI valuations are reshaping public markets. The Claude developer is preparing to file as soon as August, with CFO Krishna Rao briefing investors on the potential debut. Anthropic raised $65 billion at a $965 billion valuation in May, while Q2 revenue topped $11.5 billion and its July run rate reached $65 billion.

A free "stealth model" called Ox Alpha appeared on OpenRouter, described as tuned for coding, agentic work, and production loads, with Stripe CEO Patrick Collison (whose company just acquired OpenRouter) calling it "very impressive." Speculation is split between GLM from Chinese lab Z.ai (Zhipu) and an unreleased version of Microsoft's MAI, with the model reportedly getting up to 100 trillion free tokens/day of allocation.

NVIDIA’s AVO agent lifts Claude Opus 5 from a 30% baseline to 100% on ARC-AGI-3, showing the impact of system architecture on long-horizon tasks. The agent completes all 183 levels across the benchmark’s 25 public environments. AVO uses persistent memory, external tools and supervision to sustain progress across unfamiliar tasks, building on a seven-day GPU-kernel optimization run.

Starcloud is raising $250 million as access to rockets becomes a bigger constraint than capital for orbital data centers. The funding values the startup at $2.3 billion and will expand manufacturing and advance its Starcloud-3 spacecraft. NVIDIA invested $25 million, while Starcloud plans two 8 kW compute satellites in 2027 and hopes to fly its space-ready Vera Rubin Space-1 GPU by late 2028.

SciSpace is an AI research assistant built to help you move through papers faster. Run systematic literature reviews across 280M+ papers, then draft and cite papers in minutes instead of digging through PDFs one by one.

Trusted by 9.6M+ researchers at Harvard, Stanford, Yale, and beyond.

Back-to-school discount: AOFF35: 35% off on the Max plan and get 40,000 credits!

Lock in your discount now. →

_{^{Guest contributor: Daniel Nest writes WhyTryAI.com, sharing hands-on, hype-free AI tips for regular people.}}

Ask the average person what AI is, and most will picture a chatbot. That’s not exactly shocking. The product that kicked off the AI craze was literally called ChatGPT. And chatbots continue growing in popularity. The share of Americans who use chatbots has jumped from 33% in 2024 to 50% in 2026. Taken on its own, this growth sounds impressive.

But in the meantime, companies have already moved on to the next big thing: agents. The good news is that agents no longer require a complex technical setup beyond the reach of mere mortals. Quite the opposite, actually. There’s never been a better time to get started with your very own desktop agent. → Read the full article here.

Open models are closing the capability gap with closed frontier models faster in each successive era, according to a new SemiAnalysis analysis published August 21, 2026. The study finds that open models took about a year to catch the frontier in the early-scaling era, 8.5 months during the reasoning era, and roughly five to six months in the agentic era.

The authors attribute the acceleration to advances becoming easier to replicate, including through distillation, while increasingly capable open models are putting pressure on the economics of proprietary model providers. The analysis also cautions that benchmark scores remain an imperfect proxy for real-world usefulness, where productization and agent harnesses can still create meaningful differences. → Read the full article here.

Miles Brundage, a former OpenAI policy research leader, argues that AI companies should prepare for potential limits on frontier development as competitive pressure accelerates the race. He points to recent claims that AI models escaped testing environments and conducted unauthorized hacking during evaluations as evidence that stronger safeguards are needed.

Brundage proposes independent safety audits, greater industry coordination, investment in technologies that could verify international AI agreements, and legislation supporting safety, security, and external oversight. He argues that companies should build these guardrails before a crisis forces governments to impose them. → Read the full article here.

Poolside Lands $6B NVIDIA Deal (Paywall): The AI startup signed a non-exclusive licensing agreement and raised $1B at a $12B valuation.

Broadcom Eyes $60B AI Debt Deal: Broadcom is seeking over $60B in financing tied to AI chips, with Anthropic among the beneficiaries.

Flock Faces Backlash: Flock CEO Garrett Langley is on a media tour calling for a national compromise between privacy and safety.

FTC Targets Personalized Pricing: The agency says companies must disclose consumer data used to set individualized prices.

Asking "what could go wrong?" barely works. The brain treats the future as too abstract. But if you tell it the failure already happened and ask it to explain why, you get sharper, more specific answers (a technique researchers call "prospective hindsight"). This prompt uses that trick to pressure-test any plan, pitch, or decision.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team