Why Anthropic believes its latest model is too dangerous to release

Anthropic is delaying general release of its new Claude Mythos Preview LLM after internal tests found it discovered thousands of high-severity vulnerabilities (99% unpatched) and achieved a 72% success rate producing a Firefox exploit versus <1% for its previous model. Access is limited to ~50 entities (11 coordinating via Project Glasswing, including Google, Microsoft, Nvidia, Amazon, Apple) and Anthropic is donating $100M in access credits; access pricing is $25 per million input tokens and $125 per million output tokens. Anthropic cites safety risks, compute constraints (model is very expensive to serve) and strategic incentives to limit distribution; the firm also reports revenue run-rate growth (to $30B annualized) amid surging demand.

Analysis

Deployments of frontier-capable models will bifurcate the AI market into (a) tightly controlled, high-ARPU managed offerings sold through cloud/security channels and (b) commoditized, lower-ARPU on-device/smaller-model alternatives. Over the next 12–24 months expect cloud providers to extract premium pricing for “hardened access” and monitoring, creating a durable incremental margin line that is underpriced into current consensus for incumbent hyperscalers.

Compute scarcity and gated rollouts create a transitory supply advantage for GPU suppliers and cloud hosts, but also raise the probability of demand rationing and higher effective unit prices that depress gross margins for vertically integrated users (e.g., retail/fulfillment heavy operators) in the near term. A 3–9 month capacity shock could materially lift NVDA’s revenue per GPU while simultaneously creating a 6–18 month headwind to companies forced to buy or overprovision cloud instances for safe/monitored model use.

Regulatory and reputational catalysts are asymmetric and fast: a high-profile exploit or leak would trigger immediate usage curbs, enterprise purchasing freezes, and accelerate government procurement of “safe” enterprise LLMs — shifting procurement toward vetted cloud vendors and security-first providers. The biggest tail risk that reverses winners is rapid model replication or open-source parity; if frontier capabilities leak wholesale, the premium for managed access collapses within weeks and benefits open-source tooling and low-cost compute arbitrageurs.

Key Decisions for Investors

• Long MSFT (12-month): buy MSFT Jan-2027 1.0x ITM calls (or equivalent notional via call spread). Rationale: biggest beneficiary from premium managed LLM + security stack. Position size: 1.0–1.5% AUM. Reward: capture cloud ARPU re-rating; Risk: regulatory or multi-cloud deal losses. Stop-loss: 25% of option premium.
• Long GOOGL (12-month) vs short AMZN (12-month) pair: buy GOOGL Jan-2027 calls and buy-protective puts or sell AMZN Jan-2027 calls to fund. Rationale: Google Cloud and security tooling best-positioned to monetize gated LLM access; Amazon faces margin pressure from higher raw compute costs and uneven retail leverage. Target: asymmetric 2:1 upside vs downside over 12 months. Size: 1.0% net long exposure.
• Directional NVDA (3–6 month options spread): buy NVDA 3–6m call spread (long ~6% OTM, short ~18% OTM) to capture sustained GPU tightness while limiting premium spend. Rationale: near-term supply tightness + pricing power. Size: 0.5–1.0% AUM. Reward: high single-to-double digit return if procurement keeps tight; Risk: capacity relief or aggressive capex by competitors compresses upside.
• Hedge/insurance (6–12 months): buy a small basket of deep-OTM protection on enterprise SaaS names or buy a cybersecurity ETF/puts (size 0.5% AUM) to guard against a reputational/regulatory shock that freezes enterprise AI spending. Rationale: exploit asymmetry—large market moves on a narrow catalyst. Reward: non-linear protection; Cost: small premium drag.