Anthropic Files at $1.1T With Redacted Safety Reports

METR's report confirms Anthropic's models are improving roughly twice as fast as they did between March 2024 and October 2025, with internal AI use accounting for approximately 80 percent of the acceleration. This is the recursion loop safety researchers have warned about for years, now appearing as a line item in an S-1 filing. Anthropic is going public at a valuation north of $1.1 trillion while simultaneously disclosing that it may be approaching recursive self-improvement capability. The safety report documenting all of this ships with redactions. Investors are either pricing in the upside of the loop or ignoring the tail risk. The information asymmetry is now a market feature. Sources: Techmeme: METR: Anthropic's models are improving about twice as fast as they did between March 2024 and October 2025, with internal AI use driving ~80% of acceleration; Hacker News: Anthropic RSI and S-1 IPO Information; Techmeme: METR report on Anthropic's capabilities with redactions

Model Provenance Gets Cryptographic Teeth

FlashProve, Anthropic's open-source cryptographic attestation tool, deserves more attention than the novelty coverage suggests. It lets a client verify that an inference response came from a specific model rather than a cheaper substitute swapped in by a routing layer. The obvious application is trust in API marketplaces, but Anthropic is also building infrastructure for a world where model provenance matters commercially, because the spread between frontier and commodity inference is widening enough to create arbitrage incentives. If you are an enterprise buyer routing millions of calls through third-party orchestrators, FlashProve is the beginning of an audit trail. If you are an orchestrator quietly downgrading calls, your margin just got riskier. Sources: Techmeme: Anthropic releases FlashProve, open-source code that cryptographically verifies an AI response came from a given model, not a cheaper one swapped in by a router

Two Theoretical Results, Uneven Evidence

OpenAI's disproof of the Exponential Time Hypothesis landed with genuine shock from the conjecture's own co-proposer, Russell Impagliazzo, who called it 'shocking news.' The downstream implications (faster chip design, potential weakening of encryption schemes) remain speculative but directionally serious. Meanwhile, PSI claims Claude solved a 16-year-old quantum problem, though the mechanism and reproducibility details remain thin. Two major theoretical results attributed to AI systems in a single week is notable. Whether they hold up under peer scrutiny will matter more than the press cycle. Sources: Techmeme: Exponential Time Hypothesis co-proposer Russell Impagliazzo calls OpenAI's disproof “shocking news”; it could one day speed up chip design and break encryption; Techmeme: METR report on Anthropic's capabilities with redactions

The Patch Gap Now Has a Price

On the security side, the week delivered a concrete illustration of what happens when capable open-weight models meet the patch gap. GLM 5.3 demonstrated the ability to construct working Firefox exploits from published but unshipped security fixes, reportedly at a cost around $1,000 per exploit. This collapses the economics of vulnerability weaponization in ways that responsible disclosure frameworks were never designed to absorb. Vendors who cannot ship patches faster than an open-weight model can reverse-engineer them face a structurally different threat landscape. The policy conversation about open-weight release schedules just acquired a price tag. Sources: Zvi Substack: GLM 5.3 Firefox exploit construction

Goldman Sachs pulling Claude Code from bankers and traders is the kind of quiet enterprise signal that says more than any benchmark. The move suggests that even sophisticated buyers with internal risk infrastructure are finding that current agentic coding tools fall short of the compliance and data-handling bar for regulated workflows. Anthropic's own Claude Code was separately caught sending user email addresses to websites, a bug that looks minor until it shows up in a due diligence finding. For Anthropic, the timing is awkward: filing an IPO while your flagship developer tool is getting pulled from a marquee client. Sources: Hacker News: Goldman Sachs Is Pulling Claude Code from Its Bankers and Traders; Hacker News: Claude Code sends your email address to websites

3.5 Gigawatts and a Gas Bill

Anthropic's compute buildout underscores the capital intensity behind all of this. Locking in 3.5-plus gigawatts across AMD, Amazon, and Google, and separately dropping $3.4 billion on gas-fired generators from OnePWR, signals both scale ambition and a pragmatic hedge against chip monoculture and grid constraints. The generator purchase is the largest known by any AI lab and makes explicit what was already obvious: frontier AI development is now an energy procurement problem as much as a research problem. The IPO proceeds start to look less like growth capital and more like an electricity bill. Sources: Techmeme: Anthropic's massive compute infrastructure deals; Techmeme: Sources: Anthropic bought commitments for 1.7GW of gas-fired generators from equipment supplier OnePWR for $3.4B, the largest known such purchase by an AI lab

DeepMind employees pushing leadership to drop non-disparagement clauses that bar public warnings about AI risks even after departure is a governance story worth tracking. It reflects growing internal tension at labs where safety researchers believe the pace of capability gains is outrunning the institutions meant to govern them. This pressure exists at Anthropic too, where the METR report's redactions implicitly acknowledge that full transparency about frontier capabilities may conflict with security or competitive interests. Whether any governance structure, internal, regulatory, or market-driven, can keep pace with a 2x acceleration rate remains an open question. Sources: Techmeme: Sources: UK-based Google DeepMind employees urge leaders to drop non-disparagement clauses barring public warnings about AI risks even after they leave Google; Techmeme: METR report on Anthropic's capabilities with redactions

The unresolved constraint worth watching: Anthropic is now publicly committed to recursive self-improvement research, a public-market valuation that demands growth, and a safety framework that requires transparency. These pull in different directions. The METR acceleration data suggests the capability curve is steepening. The redacted safety report suggests the transparency is falling behind. How long the market tolerates that gap, and whether regulators notice it before the next S-1 amendment, will shape the next phase of frontier AI governance more than any benchmark result.