AI discovery: 2026/04

01.04.2026 – 30.04.2026

Product

• Anthropic releases Claude Opus 4.7 on 16 April at the same $5/$25 per million tokens as Opus 4.6, adding an xhigh effort level and gains in advanced software engineering
• OpenAI announces GPT-5.5 on 23 April for ChatGPT and Codex, with the API following at $5/$30 per million tokens, double GPT-5.4
• Cursor 3 ships on 2 April as a unified agent workspace with an Agents Window for local and cloud agents and a plugin marketplace
• Windsurf 2.0 arrives on 15 April with an Agent Command Center and one-click handoff from local Cascade sessions to Devin in the cloud

UX/UI

• Anthropic Labs launches Claude Design on 17 April, built on Opus 4.7, with export to Canva, PDF, PPTX and HTML
• GitHub Copilot for JetBrains gets inline agent mode in public preview, plus Next Edit Suggestions previews and global auto approve
• GitHub moves the global pull requests dashboard to an opt-out public preview with a reworked inbox and PR list
• Copilot Chat on github.com can now summarize, review and answer questions about a pull request or a single diff

Engineering

• Anthropic’s postmortem traces Claude Code quality complaints to three changes (lower default effort, a thinking-clearing bug, a verbosity prompt), all resolved by 20 April in v2.1.116
• Copilot cloud agent is no longer limited to pull requests: it can plan before coding, push to a branch without opening a PR and run research on a codebase
• GitHub CLI adds gh skill (v2.90.0+) to discover, install, pin and publish agent skills for Copilot, Claude Code, Cursor, Codex and Gemini CLI
• The OpenAI Codex app update adds computer use on Mac, an in-app browser, PR reviews and SSH to remote devboxes
• GitHub pauses new Copilot Pro, Pro+ and Student signups on 20 April and removes Opus models from Pro as it tightens usage limits

DevOps

• Cloudflare makes Sandboxes and Containers generally available on 13 April, with persistent agent environments and an egress proxy that injects credentials so agents never see them
• GitHub Actions OIDC tokens get an immutable sub claim built from owner and repository IDs, so a recycled repository name cannot inherit trust
• Copilot cloud agent starts over 20% faster by running on Actions custom images
• Copilot code review will consume GitHub Actions minutes on private repositories from 1 June 2026
• Dependabot-based dependency graphs for Python (pip, uv, Poetry) give fuller transitive trees and SBOMs without using Actions minutes

Infosec

• Anthropic launches Project Glasswing on 7 April, giving partners access to the unreleased Claude Mythos Preview to find and fix vulnerabilities, backed by up to $100M in usage credits
• Vercel confirms a breach that started with a compromised Context.ai tool used by an employee and exposed some customers’ non-sensitive environment variables
• OX Security reports that MCP’s STDIO transport in the official Python, TypeScript, Java and Rust SDKs executes commands by design, leading to 10 CVEs; Anthropic calls the behaviour expected
• OWASP’s Q1 2026 GenAI exploit round-up covers eight incidents, including the Flowise CVE-2025-59528 exploitation on 7 April and the LiteLLM supply-chain breach

Tricks & Tips

• Anthropic’s guide to Opus 4.7 in Claude Code: front-load the task, batch questions, keep effort at xhigh and spell out when to use subagents and tools
• A measurement of the Opus 4.7 tokenizer finds about 1.3x more tokens on real Claude Code content, roughly 20-30% higher session cost at unchanged prices
• Guide to running Gemma 4 26B-A4B locally with LM Studio’s headless lms CLI and pointing Claude Code at it through its Anthropic-compatible endpoint

Reddit & Twitter

• Hacker News thread on the Opus 4.7 announcement, where commenters argue about adaptive thinking, hidden reasoning and perceived quality regressions
• Hacker News thread on Project Glasswing, split between doubt about vendor claims and credit for AI-found vulnerability reports
• Hacker News thread on the Claude Code postmortem, focused on silent cache and thinking changes and requests for visible controls
• Simon Willison’s blog: a quantized Qwen3.6-35B-A3B on his laptop draws a better pelican on a bicycle than Opus 4.7, though he doubts it is the more capable model

OSS

• Google releases Gemma 4 on 2 April under Apache 2.0 in E2B, E4B, 26B MoE and 31B dense sizes with up to 256K context
• DeepSeek releases V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active) on 24 April as a preview with open weights and 1M context
• Alibaba releases Qwen3.6-27B, a dense Apache 2.0 model scoring 77.2 on SWE-bench Verified
• Moonshot open-sources Kimi K2.6, a 1T-parameter MoE with 32B active, 256K context and 58.6 on SWE-Bench Pro