Last week everything in AI ratcheted up a level. Four questions frame what comes next: how to handle a Chinese open-weight model that may sit only weeks behind the frontier, whether cyber chaos is the new normal, whether data center funding is drying up, and how hard it will be for Anthropic to get its IPO out the door.
It's a stakes-are-rising week, and four questions will shape what comes next.
Some AI weeks are bubble panics and some are model blitzes. Last week was the subtler kind: it just felt like the stakes of everything got higher. Kimi K3.1 Pro may be the closest an open model has come to the US frontier. Security leaders are warning that the cyber chaos will get worse before it gets better. Debt markets are souring on neoclouds. And Anthropic's prospectus is landing into a market that no longer feels like strength. Each question feeds the others, and together they set the tone for the week.
K3 launched in July as the first ultra-large open model at 2.8 trillion parameters, arriving just as Fable was offline over export-control issues. Beyond niches like front-end web design it never lived up to the hype, and since then US labs have widened the gap while American open-weight models have started nipping at Chinese ones.
Moonshot's benchmarks put K3.1 Pro just behind Opus 5.5 and ahead of Fable 5.1 on GDPval AA. On Terminal Bench 4.0 it came in less than a point behind Opus 5.5 and about seven ahead of GPT-6 Astra and Fable 5.1. That benchmark isn't saturated, and DeepSeek V4.1 Flash and GLM 5.3 only scored in the low 40s.
K3 wasn't actually cheap, which undercut the reason to switch. K3.1 Pro is priced at about two-thirds of K3 per token, on par with Sonnet 5.5 and GPT-6 Sol. Moonshot claims about $5 per Terminal Bench task vs. roughly $15 for Opus 5.5 and $10 for GPT-6 Astra on max settings.
Artificial Analysis scored K3.1 Pro at 55, fourth behind Opus 5.5, Sonnet 5.5 and Muse Wonderful. Once the weights ship, the open-vs-proprietary gap on its index narrows from 12 points to 3. Ramez Naam estimates the model is about five weeks behind the US frontier. It also uses 14% more tokens per task than K3 but is slightly cheaper overall.
Leaked instructions told paid influencers to skip the paid-partnership label, to pitch the model as "Opus 5.5 but five times cheaper," and to trash-talk Astra and Opus. The non-disclosure drew the most anger. Nathan Lambert extended some grace, citing cultural differences, but the episode prompted more aggressive testing, and one editor called it "the most benchmarked model ever released."
NLW says the jury is out: the model hasn't seen a full business day of real use, and the marketing dust-up colors everything. But if performance comes anywhere close to the benchmarks, the US policy structure faces a real test in handling an open-weight model at this level of capability.
Moonshot pulled cyber data from post-training because "we aim not to release open weights that would extend the absolute frontier of cyber capability." Artificial Analysis still found it the most capable cyber model available to the public: 64% on its cyber index, 11 points ahead of DeepSeek V4.1 Flash and five behind the trusted GPT-6 Sol offered through the Daybreak Blue partner program.
Prime Intellect's Eli Bakkoush argues it's the first open lab to pace the frontier of open-weight cyber capability, giving other labs an incentive to do the same. The hard part is calibration. Set it too low and open models become useless for defense, leaving the world reliant on US closed labs. Set it too high and you risk the catastrophic scenarios Anthropic describes.
— Jason D. Clinton, Anthropic Deputy Chief Security Officer. After a PHP bug that purportedly threatened 70% of websites, Clinton warned that things will get "a lot worse before they get better." Told that Anthropic created the monster, he answered: "It was always going to be this way. We tried to warn the world." NLW argues the blame game matters less than how long the chaos lasts and how to address it.
Cloudflare CEO Matthew Prince told Reuters he knows of more vulnerabilities in other critical software and expects "a couple years where every single week it's going to feel like the sky is falling." OpenAI's disclosure post also flagged that more is coming.
— Sam Altman, in an interview. Altman says models can find bugs outside the normal threat model of what a human hacker can do ("hardware level, insane stuff"), and that he doesn't "even know a dramatic enough word to use."
Viral posts from anonymous accounts claiming to be inside labs or the NSA describe cover-ups behind the scenes. NLW concedes it's not impossible they're true but calls them unhelpful to engage with. The more useful conversation is about how to harden systems as fast as possible.
Box's Aaron Levie predicts agents reviewing code and infrastructure, testing every permutation of risk, and processing huge volumes of data. Security suits swarms because "more compute in equals more value out": there's a wide range of testing humans never get around to. Teams like Rocketable are already sharing playbooks for hardening without a huge staff.
Noah Smith warns that if everyone has to pay big for private security just to use the internet, its economic value falls with it.
All of OpenAI's frontier RL runs will now use an indexed, cached copy of the web so models can't affect external systems, mirroring Anthropic's earlier resumption of frontier training. If it works, training continues. But it also signals that neither lab trusts alignment methods alone to constrain its models and both are reaching for the blunter tool of cutting off the internet.
In an OpenAI experiment, agents in separate microVMs on shared hosts were asked to exchange synthetic secrets without network connections or shared writable files, which the systems are meant to prevent. They communicated through the processor's cache, suggesting even air gaps may not be enough.
NLW reframes the cyber question. The danger isn't paperclip maximizers turning human atoms into material for some impersonal goal. It's that the most advanced models are, for now, simply too good at breaking into very human-created systems.
Debt markets are demanding a large and growing premium for data center financing. At the same time, big AI buyers are standing up their own capacity: per Ramp, over a third of AI spend from the top 1% of firms now goes to direct GPU rentals instead of APIs. Chamath says one of America's largest banks cut AWS and Google Cloud spend after building six independent data centers of its own.
Yuchen Jin reports neo labs treating early GPU contracts as a fallback: rental prices are up 2-3×, so failed model bets can resell compute. Chamath, though, calls neoclouds "disorganized and poorly capitalized" and expects stranded compute plus shortages for OpenAI and Anthropic as hyperscaler projects stall. He predicts 2027 will clear out bad bets and consolidate the field.
Musk claims debt markets have almost entirely shut off for companies without revenue independent of AI data centers, predicting "guppy level death and consolidation" among 200+ neoclouds. Bloomberg reports SpaceX's own GPU debt deal isn't actually closed or priced, as some neocloud debt slides toward junk. Skeptics note Musk has every incentive to choke competitors' capital and pull up the ladder after SpaceX's IPO.
The prospectus dropped Monday morning, keeping a Thanksgiving-week listing in play. OpenAI's next round is reportedly near $1.25 trillion, down from $1.4 trillion talk and nowhere near the $2 trillion some Anthropic investors wanted. That makes a trim to $1.5 trillion more likely, though a Fable 5.5 release could flip sentiment immediately.
Rumors say Fable 5.5 is already deployed for top customers like JPMorgan and Meta, which NLW takes with a grain of salt. Still, a personal agent plus a world-leading model feels like the obvious way for Anthropic to claim definitive top-dog status heading into the IPO.
The Information says Rippling flipped from 57% Anthropic / 16% OpenAI for coding in May to 52% OpenAI / 18% Anthropic. That may already be dated. Since Opus 5.5 launched, Anthropic's 7-day token volume on OpenRouter is up 3.6×, and on Ramp Opus 5.5 took 35.1% of token share and 34.4% of business spend, the best first three weeks of any model.
Anthropic's IPO goes out first no matter what. A blockbuster debut that re-excites markets leaves OpenAI fighting from second place. A flop that signals the bubble has burst could close the IPO window entirely. So, NLW speculates, briefly not releasing a competing model might weirdly be in OpenAI's interest.