Version 2

Anthropic's 2026 frontier model releases form a four-model sequence: Claude Fable 5 and Claude Mythos 5 (June 9), Claude Sonnet 5 (June 30), and Claude Opus 5 (July 24) [1][4][5]. Fable 5 is the publicly available frontier model; Mythos 5 is the same underlying model with cybersecurity safeguards lifted, restricted to vetted partners under a program called Project Glasswing that required US government approval and carries citizenship-based access restrictions [1][2][3]. Opus 5 leads the Artificial Analysis third-party leaderboard and scores three times higher than competitors on ARC-AGI 3, though Ars Technica characterizes it as an incremental token-efficiency improvement rather than a capability breakthrough [6][9]. Anthropic's claim that Opus 5 is its 'most aligned model to date' is disputed by commentators who argue the company is conflating benchmark optimization with actual alignment [8].

The Fable 5 / Mythos 5 dual-tier structure — one model publicly available, an identical model with lifted safeguards available only to government-vetted partners — concentrates access to the most capable AI systems in a small, state-approved set of actors, raising questions about who controls frontier capabilities and on what basis. Opus 5's reduction of prompt-injection attack success rates from roughly 7% to under 1% matters practically for agentic deployments, since prompt injection is the primary attack surface in autonomous multi-step tasks.

Anthropic's mid-2026 model sequence begins on June 9, 2026, with the simultaneous release of Claude Fable 5 and Claude Mythos 5 [1]. Fable 5 is the publicly available flagship, described as state-of-the-art across software engineering, vision, knowledge work, and scientific research. It includes safety classifiers that fall back to Claude Opus 4.8 for queries involving cybersecurity, biology, and model distillation, triggering in under 5% of sessions [1]. Mythos 5 is the same underlying model with those classifiers removed, available through a restricted program called Project Glasswing for vetted partners and select biomedical researchers [1]. Mythos 5's autonomous capabilities include week-long scientific research tasks: in one documented case it trained a genomics model that outperformed a recently published Science journal paper while being 100 times smaller [1]. Both models are priced at $10 per million input tokens and $50 per million output tokens [1]. Access to Mythos 5 required US government approval — the Trump administration formally cleared the commercial release in late June 2026 — and carries citizenship-based restrictions that generated concern among non-US developers about unequal access to frontier capabilities [2][3]. Anthropic requires 30-day data retention for all Mythos-class traffic [1].

Claude Sonnet 5 followed on June 30, 2026, positioned as an agentic model approaching Opus 4.8 performance at a lower price tier, becoming the default for Free and Pro consumer plans [4]. Claude Opus 5 arrived July 24, 2026, with Anthropic claiming leadership on Frontier-Bench v0.1, an ARC-AGI 3 score three times higher than the next-best model, and per-task performance more than double Opus 4.8's at identical list pricing [5]. A third-party data point supports part of this: Opus 5 led the Artificial Analysis leaderboard at launch, placing above Fable 5 [6]. One improvement that received less initial attention than the benchmark claims is prompt-injection resistance: the Opus 5 system card documents a reduction in computer-use attack success from roughly 7% to under 1%, which Boris Cherny described as 'a bit buried' at page 73 of the card [7][8]. Zvi Mowshowitz confirms this as a practical advance — 'not getting hijacked via prompt injection is the key to unlocking the confidence to do a host of activities you otherwise can't do' — and separately notes Opus 5 now permits source-code vulnerability discovery at all access levels while continuing to block vulnerability discovery in compiled binaries [8].

External reactions split on framing. Ars Technica's Samuel Axon characterizes Opus 5 as a noteworthy but incremental update centered on token efficiency, not a capability breakthrough on the scale of Opus 4.5, though he notes it has been popular for coding tasks [9]. Zvi Mowshowitz accepts the practical safety and capability improvements as genuine but strongly disputes the 'most aligned model to date' designation, arguing that optimizing against alignment metrics and achieving actual alignment are distinct things, and that conflating them is 'one of the most irresponsible and destructive possible mistakes Anthropic can make in their position' [8]. Simon Willison highlights Opus 5's agentic proactivity — when given a geometry problem without a tool, the model wrote its own computer vision pipeline from raw pixels — and describes Anthropic's framing of Opus 5 as approaching Fable 5's intelligence at half the price as promising [6].