AI Industry
News Digest

GPT-5.4-Cyber Opens Vetted Access for Defensive Security Teams

OpenAI launched GPT-5.4-Cyber on 14 April, a fine-tuned variant of GPT-5.4 built specifically for defensive cybersecurity work. If you work in security research or run a vetted security team, this model is worth paying attention to: it has a lower refusal threshold than standard OpenAI deployments and adds binary reverse engineering capabilities, meaning it can analyse compiled software for vulnerabilities without requiring access to source code.

Access is deliberately restricted. Individuals can verify their identity at chatgpt.com/cyber, while enterprise teams need to request access through an OpenAI representative. OpenAI began its Trusted Access for Cyber (TAC) programme in February 2026, and this release expands it significantly. The announcement follows Anthropic's Project Glasswing earlier in April, signalling that purpose-built cybersecurity models are now a distinct product category rather than an edge use case.

Claude Code Gains Scheduled Routines and a Redesigned Desktop Interface

Anthropic launched two notable updates to Claude Code on 14 April: a new routines feature and a redesigned desktop app. Routines let you package automations and set them to run on a schedule or on a trigger. Crucially, they run on Anthropic's own infrastructure, so your computer does not need to be online for each task. Pro users receive five routines per day, Max users fifteen, and Team and Enterprise users twenty-five. If you have been using Claude Code for repetitive tasks such as scheduled GitHub actions or API workflows, this removes a key friction point.

Alongside routines, the desktop app gained a multi-session sidebar to run several conversations from one window, an integrated terminal, in-app file editing, HTML and PDF preview, and a drag-and-drop layout you can rearrange to suit your workflow. The redesign is available to Pro, Max, Team, and Enterprise users on macOS and Windows.

Anthropic Adjusts Claude Enterprise Pricing to $20 Per User Per Month

Anthropic confirmed on 15 April that its Claude Enterprise Edition will move to a per-user pricing model at $20 per user per month. If your organisation has been evaluating Claude Enterprise, this gives you a clearer cost baseline than previous arrangements. The change reflects a broader shift in how frontier AI providers are structuring commercial access: moving away from usage-only billing towards predictable seat-based pricing more familiar to enterprise software buyers. The structure mirrors how Microsoft 365 Copilot and similar tools are priced, which may simplify procurement conversations for teams running multi-vendor AI evaluations.

Anthropic Set to Launch Claude Opus 4.7 and an AI Design Tool This Week

The Information reported on 14 April, citing sources with knowledge of Anthropic's plans, that the company is preparing to release Claude Opus 4.7 alongside a new AI-powered design tool for websites, presentations, and product mockups as soon as this week. Opus 4.7 would be an incremental upgrade from Claude Opus 4.6, released in February 2026.

The design tool angle attracted significant market attention: shares in Figma, Adobe, Wix, and GoDaddy each fell 2 to 4 per cent on 14 April following the report, as investors assessed competitive risk from a well-resourced entrant into AI-assisted design. Anthropic has not yet made a public statement confirming or denying the launch timing, so treat this as a highly credible signal of what is coming rather than an official announcement.

NVIDIA Ising: World's First Open-Source AI Models for Quantum Computing

NVIDIA released the Ising model family on 14 April, described as the world's first open-source AI models designed specifically to accelerate quantum computing. The models target two of the most critical bottlenecks in building fault-tolerant quantum processors: calibration and error correction. Ising delivers up to 2.5 times faster performance and 3 times higher accuracy compared with existing open-source approaches for these tasks.

For anyone building on or evaluating quantum computing infrastructure, this is a meaningful practical tool rather than a research curiosity. Early adopters already include Harvard John A. Paulson School of Engineering and Applied Sciences, Fermi National Accelerator Laboratory, and the UK National Physical Laboratory, lending credibility to the performance claims.

RTX PRO 5000 72GB Blackwell Workstation GPU Now Generally Available

NVIDIA's RTX PRO 5000 72GB Blackwell GPU became generally available on 15 April through distribution partners including Ingram Micro, Leadtek, Unisplendour, and xFusion. The 72GB memory configuration extends the RTX PRO 5000 workstation line and is aimed squarely at local AI inference and agentic workloads that demand large on-device memory capacity. If you are speccing AI-capable workstations for a team or evaluating hardware options for running large models locally, this adds a commercially available option with substantial VRAM headroom.

Agent Bricks Document Intelligence and Custom Agents Reach General Availability

Databricks moved Document Intelligence and Custom Agents to general availability on 14 April, alongside expanded AI Gateway capabilities across the Agent Bricks platform. If you are building enterprise AI applications on Databricks, the GA designation removes the preview-stage caveats and signals that the platform is considered mature enough for production workloads. The new AI Gateway additions expand the governance and security controls available to teams managing multiple agents across an organisation, which has been a practical blocker for enterprise adoption.

This positions Agent Bricks as a governed enterprise agent platform rather than an experimental tool, with the infrastructure-level data access Databricks customers already rely on sitting underneath it. The timing puts Databricks directly in competition with Anthropic's Managed Agents and Amazon Bedrock AgentCore as the enterprise agent hosting market takes shape.