Version 3

Anthropic's 2026 model sequence — Fable 5 and Mythos 5 (June 9), Sonnet 5 (June 30), Opus 5 (July 24) — now centers on an unresolved regulatory episode: multiple news reports indicate the US government suspended access to both Fable 5 and Mythos 5 within days of their commercial rollout, reversing an earlier approval, while Cisco and Dragos reportedly retained access under separate conditions [4][5][6]. Before the suspension, Project Glasswing partners had identified more than 10,000 high- or critical-severity security flaws using Claude Mythos Preview, and the program had expanded to roughly 200 organizations across 15+ countries [2]. Opus 5, released July 24 at $5 per million input and $25 per million output tokens, is positioned as near-Fable-5 performance at approximately half the cost [12], with benchmark leadership claimed on Frontier-Bench v0.1 and ARC-AGI 3 [11] and a 'most aligned' designation that external analysts dispute [15].

A government that approves then suspends frontier AI model access within days shows the regulatory environment for dual-use cybersecurity AI is unstable. Anthropic's proactive-defense rationale for Project Glasswing — that Mythos-class capabilities will reach other labs within 6-12 months regardless of what Anthropic does [2] — now operates against active government restriction rather than in a permissive environment, and the terms on which specific partners retain access are not publicly explained.

Anthropic released Claude Fable 5 and Claude Mythos 5 simultaneously on June 9, 2026 [1]. Fable 5 is the publicly available flagship, with safety classifiers that fall back to Claude Opus 4.8 for queries involving cybersecurity, biology, and model distillation, triggering in under 5% of sessions. Mythos 5 is the same underlying model with those classifiers removed, distributed through Project Glasswing — a restricted program for vetted partners and select biomedical researchers, priced at $10/$50 per million tokens. As of June 2, 2026, before the full launch, Glasswing partners using Claude Mythos Preview had already identified more than 10,000 high- or critical-severity security flaws, and Anthropic announced expansion to approximately 150 new organizations, reaching about 200 total partners across power, water, healthcare, communications, and hardware sectors in 15+ countries [2]. Anthropic's stated rationale is explicitly preemptive: it expects other AI companies to reach Mythos-class capability within 6-12 months and argues that safeguarded deployment for defense now is preferable to waiting for uncontrolled proliferation [2].

The regulatory picture then shifted. The Trump administration cleared Mythos 5 for commercial access in late June 2026 [3], but multiple reports indicate the US government subsequently forced Anthropic to pull access to both Fable 5 and Mythos 5 within days of their broader commercial rollout [4][5]. Cisco and Dragos reportedly retained Mythos access after the suspension order [6], and news accounts frame the episode in geopolitical terms [7][8]. The criteria distinguishing which partners retain access have not been publicly explained. Citizenship-based restrictions already embedded in the Mythos 5 rollout had drawn objections from non-US developers who saw no clear path to equivalent access [9].

Claude Sonnet 5 followed on June 30, 2026, as an agentic model approaching Opus 4.8 performance at a lower price, becoming the default for Free and Pro consumer tiers [10]. Claude Opus 5 arrived July 24, priced at $5 per million input tokens and $25 per million output tokens — roughly half Fable 5's cost — with Anthropic claiming benchmark leadership on Frontier-Bench v0.1 and an ARC-AGI 3 score three times higher than the next-best model [11][12]. Third-party data partially supports this positioning: Opus 5 led the Artificial Analysis leaderboard at launch, placing above Fable 5 [13]. One security property received less attention than the benchmark claims: the Opus 5 system card documents a reduction in computer-use prompt-injection attack success from roughly 7% to under 1%, which Boris Cherny noted was buried at page 73 of the card [14][15]. The Neuron separately reported that Anthropic and AMD agreed to deploy up to 2 GW of AMD accelerators for Claude, and that Fable 5 reportedly produced a short counterexample to the 87-year-old Jacobian conjecture [12].

External reactions divide on framing. Ars Technica characterizes Opus 5 as an incremental token-efficiency improvement, not a capability breakthrough [16]. Zvi Mowshowitz accepts the practical safety improvements as genuine but disputes the 'most aligned model to date' designation, arguing it conflates optimizing against alignment benchmarks with achieving actual alignment [15]. Ethan Mollick, writing July 23, categorizes Claude and ChatGPT as the only viable general-purpose agentic platforms at the $20/month consumer tier and describes prompt injection as an unresolved practical risk for autonomous deployments — a concern the Opus 5 system card data addresses but does not eliminate [17]. Simon Willison flags Opus 5's agentic proactivity as notable: given a geometry problem without a tool, the model wrote its own computer vision pipeline from raw pixels [13].