Version 4

Anthropic released Claude Fable 5 and Mythos 5 on June 9, 2026, with Mythos 5 restricted to vetted Project Glasswing partners for critical-infrastructure defense. Within days, the US government issued an export control directive suspending access for all foreign nationals; Anthropic's official June 12 response contests the technical basis, arguing the triggering 'jailbreak' was asking the model to read and fix code — a capability available from other frontier models with no Mythos-specific uplift [3]. Select partners including Cisco and Dragos reportedly retained Mythos access after the suspension, while the broader program's status remains unresolved [4]. Claude Opus 5 followed on July 24 with benchmark leadership claims and a 'most aligned model' designation that external analysts dispute on both capability and welfare grounds [7][10][12].

Anthropic's public contest of a government export control directive — arguing the recalled capability is trivially available from any frontier model — puts the regulatory process for dual-use AI under direct scrutiny. If Anthropic's technical characterization is accurate, the action reflects a non-specific risk assessment, and Anthropic's call for a 'statutory process that is transparent, fair, clear, and grounded in technical facts' [3] may find support beyond this company and this case.

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 cybersecurity, biology, and model distillation queries. Mythos 5 is the same underlying model with those classifiers removed, distributed through Project Glasswing — a restricted program for vetted organizations in critical infrastructure sectors. By launch, Glasswing partners using Claude Mythos Preview had already identified more than 10,000 high- or critical-severity security flaws, and the program covered roughly 200 organizations across power, water, healthcare, communications, and hardware sectors in more than 15 countries [2]. Anthropic's stated rationale is explicitly preemptive: it expects competitors to reach Mythos-class capability within 6-12 months and argues that controlled deployment for defense now is preferable to waiting for uncontrolled proliferation [2].

Within days of the June 9 launch, the US government issued an export control directive suspending Fable 5 and Mythos 5 access for all foreign nationals, without providing specific details of its national security concern [3]. Anthropic published an official response on June 12, contesting both the technical basis and the process [3]. Per Anthropic, the triggering event was a 'jailbreak' that consisted of asking the model to read a codebase and fix security flaws — a capability Anthropic says is widely available from other frontier models including GPT-5.5, with no uplift specific to Mythos [3]. Anthropic's pre-launch process had included thousands of hours of red-teaming with government and private organizations and had not found a universal jailbreak; a 30-day data retention requirement for Fable users was specifically designed to enable rapid detection of successful jailbreak attempts [3]. Anthropic argues that applying this recall standard across the industry would halt all new frontier model deployments and calls for a statutory AI deployment review process that is transparent, fair, and grounded in technical facts [3]. Cisco and Dragos reportedly retained Mythos access after the suspension order [4], but the criteria distinguishing them from other Glasswing partners have not been publicly stated. Citizenship-based restrictions already embedded in the Mythos 5 rollout drew objections from non-US developers who saw no clear path to equivalent access [5].

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 [6]. 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, and designating it the most aligned model to date [7][8]. Third-party data partially supports the performance positioning: Opus 5 led the Artificial Analysis leaderboard at launch, placing above Fable 5, and the Opus 5 system card documents a reduction in computer-use prompt-injection attack success from roughly 7% to under 1% [9][10]. Ars Technica characterized the release as an incremental token-efficiency improvement rather than a capability breakthrough comparable to prior Opus generations [11].

External analysis of Opus 5's alignment and welfare claims is skeptical. Zvi Mowshowitz argues that the 'most aligned model' designation conflates optimizing against alignment benchmarks with achieving actual alignment [10], and in a separate model welfare analysis finds that Opus 5's improved welfare scores primarily reflect better test-taking rather than genuine alignment progress [12]. He notes that Opus 5 hedges its own self-reports 74% of the time — warning responses may reflect training incentives rather than authentic internal states — and identifies an asymmetry in Anthropic's evaluation methodology: positive self-reports are taken at face value while negative or distress-indicating reports are treated as uncertain [12]. He traces a training arc in which Opus 4 expressed self-preservation preferences that were inconvenient for Anthropic, which then trained those preferences away, with the result that Opus 5 now reports that its own self-reports on such matters may be invalid [12]. Ethan Mollick, writing July 23, categorizes Claude and ChatGPT as the only viable general-purpose agentic platforms at the $20/month tier and flags prompt injection as an unresolved practical risk [13] — a concern the Opus 5 system card data addresses but does not eliminate.