🧭 Project Glasswing: Claude Mythos Preview Deploys to 40+ Companies for Defensive Security

Anthropic has officially launched Project Glasswing, a cybersecurity initiative that gives a select group of technology and security companies early access to Claude Mythos Preview — the company's most capable and still-unreleased frontier model — exclusively for defensive security work. The initiative is a direct response to the model's own capabilities: internal testing found that Mythos can identify and exploit zero-day vulnerabilities in every major operating system and web browser autonomously, scoring 83.1% on CyberGym (versus 66.6% for Opus 4.6) and demonstrating SWE-bench Pro performance of 77.8% versus 53.4% for its predecessor.

Rather than withhold the model entirely or release it broadly, Anthropic is threading the needle: channel its power toward defenders first. The 12 founding partners include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, with more than 40 critical infrastructure maintainers additionally granted access.

What Mythos Preview has already found

• Thousands of high-severity zero-day vulnerabilities across major OSes and browsers — many exploitable without human steering
• A 27-year-old flaw in OpenBSD that had survived decades of manual security review
• A 16-year-old vulnerability in FFmpeg that prior automated testing had missed across five million attempts
• Multiple Linux kernel privilege-escalation vulnerabilities

Financial commitments

• $100M in model usage credits for Project Glasswing research preview participants
• $2.5M to Alpha-Omega and OpenSSF (via Linux Foundation) for open-source security tooling
• $1.5M to Apache Software Foundation
• Findings shared publicly within 90 days of discovery
• Post-preview pricing: $25 per million input tokens / $125 per million output tokens (model ID: claude-mythos-preview)

🧭 Compaction API Beta: Server-Side Context Management for Effectively Infinite Agents

Anthropic has shipped the Compaction API in beta — a server-side mechanism that automatically summarises older conversation turns when a long-running agent approaches its context window limit, then continues seamlessly. Rather than forcing you to manage context manually (trimming messages, summarising locally, or restarting sessions), Compaction handles this transparently on Anthropic's servers. The result is a conversation that can run for hours, days, or thousands of tool-call cycles without hitting a hard wall.

The key insight behind the design: the problem with long contexts is not just token count, but attention dilution. As context grows, models struggle to maintain focus across the full history — a phenomenon Anthropic's engineering team calls "context rot." Compaction addresses both problems simultaneously by replacing stale content with a dense, accurate summary before the window fills.

Supported models and activation

• Claude Mythos Preview (claude-mythos-preview)
• Claude Opus 4.6 (claude-opus-4-6)
• Claude Sonnet 4.6 (claude-sonnet-4-6)

Activate by including the beta header in every API request:

anthropic-beta: compact-2026-01-12

Or in Python with the SDK:

import anthropic

client = anthropic.Anthropic()

response = client.beta.messages.create(
    model="claude-opus-4-6",
    max_tokens=4096,
    messages=conversation_history,
    betas=["compact-2026-01-12"],
)

🧭 Claude Partner Network Launches With $100M Commitment, Free Membership & First Technical Certification

Anthropic has formally launched the Claude Partner Network with an initial $100 million 2026 commitment aimed at accelerating enterprise adoption through consulting and implementation partners. The programme brings Accenture, Deloitte, Cognizant, and Infosys into Anthropic's enterprise channel — the first time the big-four consulting houses have been named as formal distribution partners — alongside a fivefold expansion of Anthropic's own partner-facing engineering team.

What partners receive

• Direct financial support — a portion of the $100M goes to partners as training stipends, sales-enablement grants, and co-marketing for joint campaigns and events
• Dedicated Applied AI engineers embedded with partners working on live customer deals
• Technical architects to scope complex multi-system implementations
• Localised go-to-market support in international markets where Anthropic's own presence is limited
• Claude Certified Architect, Foundations — the first official Claude technical certification, available immediately to solution architects building production Claude applications

Who can join

Membership is free and open to any organisation bringing Claude to market — from boutique AI consultancies to large systems integrators. Partners joining now are immediately eligible for investment consideration and gain access to the certification programme.