AI News Roundup — April 18, 2026
Anthropic ships Claude Opus 4.7 with major coding gains and launches Claude Design for visual prototyping, while its withheld Mythos model completes a 32-step simulated cyberattack and triggers an emergency government-banking sector meeting. OpenAI responds with its own security-focused GPT-5.4-Cyber model and signs a $20 billion compute deal with Cerebras.
Anthropic Releases Claude Opus 4.7 with Major Coding Gains
Anthropic shipped Claude Opus 4.7, its new flagship model, with SWE-bench Verified jumping to 87.6% from 80.8%, GPQA Diamond reaching 94.2%, and CursorBench hitting 70%. New capabilities include an xhigh effort level, Task budgets for agentic workflows, a /ultrareview multi-agent code review tool, and 2,576px vision support. The model is available on Claude.ai, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry at unchanged pricing ($5/$25 per million tokens).
Anthropic Launches Claude Design for Visual Prototyping
Alongside Opus 4.7, Anthropic released Claude Design — a tool that lets users create prototypes, slide decks, one-pagers, and other visual assets directly inside Claude. It reads a company's codebase and design files to automatically apply brand systems and style guides. Claude Design is available in research preview for Pro, Max, Team, and Enterprise subscribers.
Claude Mythos Completes 32-Step Simulated Cyberattack, Triggers Emergency Government Meeting
Anthropic's Claude Mythos Preview became the first AI model ever to complete a full 32-step simulated network attack in evaluations by the UK AI Security Institute. The achievement triggered an emergency meeting convened by Treasury Secretary Scott Bessent and Fed Chair Jerome Powell with the CEOs of five major US banks — underscoring why Anthropic has kept the model restricted to vetted cybersecurity partners only.
OpenAI Releases GPT-5.4-Cyber for Security Defenders
OpenAI launched GPT-5.4-Cyber, a fine-tuned variant of GPT-5.4 built specifically for legitimate defensive cybersecurity work. The model features lower refusal thresholds for tasks like binary reverse engineering and is being rolled out to thousands of verified security researchers. The release is widely seen as OpenAI's direct response to Anthropic's Project Glasswing initiative.
OpenAI Signs $20 Billion Compute Deal with Cerebras, Eyes Equity Stake
OpenAI agreed to spend more than $20 billion over three years on Cerebras-powered server capacity, with the arrangement potentially giving OpenAI warrants for a minority equity stake in Cerebras. Separately, Cerebras is targeting a $35 billion valuation in a planned IPO — making this one of the most consequential AI chip and compute partnerships announced this year.
Maine Moves to Ban New AI Data Centers
Maine advanced legislation that would ban new AI data center construction in the state, citing concerns about energy consumption, water usage, and grid stability. The bill represents one of the most aggressive state-level responses to AI infrastructure growth and could inspire similar measures in other energy-constrained regions across the US.
Startups Caught Selling User Data as AI Training Sets
A wave of scrutiny is hitting AI startups found to be quietly selling user-generated data to AI labs and training data brokers. The practice — typically buried in terms of service — is drawing regulatory attention and user backlash, with several companies now under investigation for potential violations of GDPR and US state privacy laws.