4 Questions for AI This Week

Moonshot’s Kimi K3.1 Pro is the first update to the K3 family, but its performance remains contested. Artificial Analysis scored it 55 on its Intelligence Index, ranking it fourth behind Opus 5.5, Sonnet 5.5, and Muse Wonderful, and two points ahead of Fable 5.1, GPT-6 Astra, and Gemini IV Argon. Moonshot’s own benchmarks put it just behind Opus 5.5 on GDPVAL AA and less than a point behind on Terminal Bench 4.0, where it reportedly beat GPT-6 Astra and Fable 5.1 by about seven points. Ramiz Nan estimated K3.1 Pro was roughly five weeks behind the US frontier, though some testers found it lacking and the speakers said its real-world performance was still uncertain.

K3.1 used 14% more tokens per task than K3 in Artificial Analysis testing, but its lower per-token price made it slightly cheaper overall. Moonshot priced it at about two-thirds of K3’s cost, matching Sonnet 5.5 and GPT-6 Sol; a Terminal Bench 4.0 task was estimated at about $5, compared with $15 for Opus 5.5 and $10 for GPT-6 Astra at maximum settings. It still cost almost ten times as much as DeepSeq V4.1 Flash and about twice as much as GLM 5.3. The launch was also overshadowed by a reported influencer campaign that discouraged paid-partnership labels and suggested calling K3.1 “Opus 5.5 but five times cheaper”; Chubby said the promotion damaged the launch, while Nathan Lambert pointed to possible cultural and communication differences.

Moonshot said it deliberately limited K3.1 Pro’s offensive cyber capability by removing cyber-related training data, while trying to preserve general coding skill and reasonable cyber performance. Artificial Analysis gave it a 64% cyber-index score, 11 percentage points above DeepSeek v4.1 Flash and five below the trusted GPT-6 Sol version available through Daybreak Blue. Eli Bakoush of Prime Intellect called K3.1 the first open lab to pace frontier open-weight cyber capabilities, but warned that setting a limit means balancing defensive usefulness against catastrophic risk. The speakers stressed that models able to find and exploit vulnerabilities already exist, making mitigation more urgent than assigning blame.

Warnings about worsening security ranged from Jason D. Clinton’s call to “hold on” for around 18 months to Cloudflare CEO Matthew Prince’s forecast of infrastructure hardening followed by a couple of years in which every week could feel like “the sky is falling.” Sam Altman said models can find bugs outside ordinary hacker threat models, including at the hardware level, and that he no longer believes every bug can simply be patched. OpenAI said Frontier reinforcement-learning runs would use quarantined, offline environments built from indexed and cached web content; Anthropic had also resumed Frontier training. Yet an OpenAI experiment found agents in separate micro-VMs could exchange synthetic secrets through the processor cache without network access or shared writable files. The speakers questioned whether internet isolation is enough and whether some frontier models could become too dangerous to release.

AI infrastructure financing looks more fragile as debt markets demand larger premiums and major customers build more capacity themselves. Ramp data showed that more than a third of AI spending by the top 1% of firms went to direct GPU rentals rather than model APIs. Chamath Palihapitiya said customers were prioritizing data sovereignty, citing a major American bank that cut AWS and Google Cloud spending while building six data centers. Elon Musk predicted substantial stranded compute and said debt funding had largely closed to companies without significant non-AI data-center revenue; he said SpaceX secured $40 billion in GPU debt financing, supported by its investment-grade status and Starlink and launch revenue. Palihapitiya predicted delays and consolidation among Neo Clouds, while Musk said more than 200 could face severe failures. They differed over whether Musk’s warnings reflected an AI bubble or an effort to restrict competitors’ access to capital.

Anthropic’s prospectus appeared the morning of recording, amid speculation about an IPO the week of Thanksgiving, but no valuation had been attributed to the company or its bankers. Reports put OpenAI’s next fundraising valuation near $1.25 trillion, below an earlier $1.4 trillion figure; some Anthropic investors reportedly sought $2 trillion, while another group urged a reduction to $1.5 trillion. The host said Anthropic no longer seemed to be heading toward an IPO from a position of strength, although a strong Fable 55 release could change sentiment. Market-share indicators were mixed: OpenAI price cuts reportedly helped it gain share, and Rippling’s coding split shifted from Anthropic at 57% and OpenAI at 16% in May to OpenAI at 52% and Anthropic at 18%. But OpenRouter reported Anthropic’s seven-day average token volume had risen 3.6 times since Opus 55 launched, while Ramp put Opus 55 at 35.1% of token share and 34.4% of business spend. The host said that momentum would need to last several more weeks to match Opus 4.6’s two-month run.