Claude Mythos Preview

Overview

Claude Mythos Preview is Anthropic's frontier-tier model, announced April 7, 2026, and made available via restricted partner access through Project Glasswing rather than general availability. Anthropic explicitly describes Mythos as a new model tier — not an Opus upgrade — that performs strongly across the board but is strikingly capable at computer security tasks. In Anthropic's head-to-head benchmarks, Mythos Preview leads Opus 4.7 on every test, with margins ranging from +6.3 points on SWE-bench Verified to +13.5 points on SWE-bench Pro and a generational lead on cybersecurity benchmarks.

Mythos's defining capability is offensive computer security. Anthropic has reportedly used Mythos Preview to identify thousands of zero-day vulnerabilities across major operating systems, browsers, and software. Publicly disclosed examples include a 27-year-old OpenBSD TCP SACK remote crash bug, a 17-year-old FreeBSD NFS remote code execution hole, a 16-year-old FFmpeg H.264 heap write, and a guest-to-host virtual machine monitor memory corruption primitive. In controlled evaluations, Mythos could execute multi-stage attacks on vulnerable networks and discover and exploit vulnerabilities autonomously — work that would take human professionals days. This capability directly motivated Project Glasswing (Anthropic's pre-release vulnerability hunting program with industry partners) and is widely cited as a driver of the White House's reported consideration of pre-release frontier-model review.

Specifications

• Developer: Anthropic
• Release Date: April 7, 2026 (restricted preview via Project Glasswing; also available in preview on Google Cloud Vertex AI)
• Type: Frontier large language model; multimodal
• Architecture: Reportedly ~10 trillion parameters (per third-party reporting; Anthropic has not officially confirmed)
• Training Hardware: Reportedly trained on NVIDIA Blackwell hardware (per third-party reporting)
• Pricing (Project Glasswing partners): $25 per million input tokens / $125 per million output tokens — 5x the price of Claude Opus 4.7
• Access: Restricted — Project Glasswing partner program (12 founding orgs + ~40 vetted critical-infrastructure operators); preview availability on Google Cloud Vertex AI for select customers

Capabilities

Cybersecurity (defining capability): State-of-the-art performance on cybersecurity benchmarks with generational margins over Opus 4.7. Demonstrated capacity to autonomously discover and exploit vulnerabilities across major OSes, browsers, and infrastructure software. Multi-stage autonomous attack execution on vulnerable networks.

General Frontier Performance: Leads Opus 4.7 on every Anthropic-published head-to-head benchmark, including:

• +6.3 points on SWE-bench Verified
• +13.5 points on SWE-bench Pro
• Generational lead on cybersecurity benchmarks

Demonstrated Vulnerability Discoveries (publicly disclosed):

• 27-year-old OpenBSD TCP SACK remote crash bug
• 17-year-old FreeBSD NFS remote code execution hole
• 16-year-old FFmpeg H.264 heap-write vulnerability
• Guest-to-host VMM memory corruption primitive

Limitations

Restricted Access: Mythos Preview is not generally available. Use is gated through Project Glasswing — an invitation-only partner program reportedly including AWS, Apple, Cisco, Google, JPMorgan Chase, and Microsoft (the 12 founding partners), plus ~40 vetted critical-infrastructure operators. Anthropic is committing up to $100M in usage credits across these partners and $4M in direct donations to open-source security organizations.

Pricing: At $25/$125 per million input/output tokens, Mythos is 5x the cost of Opus 4.7 — making it impractical for general workflows even when access is available.

Regulatory Pressure: Mythos's offensive cybersecurity capabilities are widely reported as a driver of the Trump administration's reportedly-considered executive order to establish a federal-industry working group reviewing frontier models pre-release. This makes Mythos's regulatory status forward-looking-uncertain.

Capability Risk Profile: AISI evaluations and third-party analyses (Centre for Emerging Technology and Security at Turing Institute, CrowdStrike) emphasize that the same offensive capabilities that make Mythos valuable to defenders make it dangerous if misused or leaked.

Recent Developments

• April 7, 2026 Announcement: Mythos Preview revealed as a new frontier tier distinct from Opus, with cybersecurity as the headline capability.
• Project Glasswing Launch: Concurrent launch of the partner program providing structured pre-release access to Mythos for vulnerability hunting; ~$100M in usage credits committed across 12 founding partners and ~40 critical-infrastructure operators.
• Vertex AI Preview: Google Cloud announced Mythos Preview availability on Vertex AI for select customers.
• AISI Evaluation: U.K. AI Safety Institute published an evaluation of Mythos Preview's cyber capabilities.
• Regulatory Catalyst: Reporting indicates Mythos's capabilities helped motivate the White House's consideration of pre-release federal review for frontier models.
• Opus 4.7 Differentiation: Released the same week as Opus 4.7 (which ships with cybersecurity safeguards on by default) — Anthropic's positioning is "Opus 4.7 for everyone, Mythos for vetted security work."

Last Updated

May 7, 2026