Sec & AI News — 24 April 2026
🟣 GPT-5.5 ships. Every benchmark moves. Price doubles.
OpenAI's first ground-up retrain since GPT-4.5 lands in ChatGPT and Codex. Terminal-Bench 82.7%, beating the Anthropic model too scary to release. Artificial Analysis Intelligence Index: solo first place, dethroning the three-way tie between Opus 4.7, Gemini 3.1 Pro, and GPT-5.4. API pricing: $5 input, $30 output per million — double GPT-5.4. Smarter, fewer tokens, wallet still bleeds.
🟢 ChatGPT Images 2.0 eats Nano Banana's lunch
gpt-image-2 ships with thinking, web search, 2K resolution, and dense multilingual text. LM Arena Elo: 1500 vs Nano Banana's 1271 — +242 point lead, biggest image-model jump on record. Renders working barcodes that actually scan. Draws 32×48 mazes with solutions. Generates magazine collages with Hindi, Korean, and Chinese text that doesn't melt. Image generation increasingly looks solved.
🎨 Claude Design turns prompts into prototypes. Figma stock drops 7%.
Anthropic Labs ships a visual design tool powered by Opus 4.7. Prompt in, polished mockups, slides, one-pagers out. Reads your codebase and design files, then every project uses your colors and typography. Also does After-Effects-style animations — bar charts, map zooms, title cards — in a few prompts. Pro/Max/Team/Enterprise only.
📊 Claude Cowork gets Live Artifacts. Dashboards that refresh themselves.
Connect your apps and files. Prompt a dashboard. Close it. Open it next week — it refreshes with current data. Saved in a Live Artifacts tab with version history. No manual updates, no re-prompting. Bloomberg Terminal territory, done with natural language.
🔬 DeepMind's Deep Research Max runs 160 queries while you sleep
Built on Gemini 3.1 Pro. MCP support, so it plugs into your custom data streams. Native chart and infographic generation inline. DeepSearchQA jumps from 66.1% to 93.3%. Humanity's Last Exam climbs to 54.6%. Runs in the background so you can do something else, like read an email.
🐉 Alibaba goes proprietary with Qwen 3.6 Max Preview
Qwen's flagship tier — closed weights. 256K context, preserve_thinking for multi-turn agents, OpenAI + Anthropic API compatible. Ranks first on SWE-bench Pro, Terminal-Bench 2.0, and a parade of QwenClaw/QwenWeb benchmarks. Chinese AI labs pivoting from free-for-all to paid API, one model at a time.
🌙 Kimi K2.6 open-weights beats GPT-5.4 on Humanity's Last Exam
Moonshot releases K2.6 under Modified MIT. 1T parameters, 32B active, 256K context, native multimodal. SWE-Bench Pro 58.6 — beats GPT-5.4 xhigh and Opus 4.6. HLE-Full with tools: 54.0, leading every frontier model in the comparison. Scales to 300 parallel sub-agents across 4,000 coordinated steps. Open weights. Hugging Face. Go.
🛡️ OpenAI Privacy Filter masks your PII before the cloud sees it
Open-weight, Apache 2.0, runs on-device. 1.5B total / 50M active parameters, 128K context. Detects and masks names, addresses, emails, phone numbers, URLs, dates, account numbers, secrets. F1 score 97.43%. Weights on Hugging Face and GitHub. Fine-tune it for your own PII categories. OpenAI open-sourcing a thing is still news.
🩺 ChatGPT for Clinicians. Free. Verified by NPI.
Physicians, NPs, PAs, pharmacists in the US. Free. Verification via National Provider Identifier. Clinical search with citations, reusable skills, optional HIPAA BAA. Conversations not used for training. Physician advisors rated 99.6% of responses safe across ~7,000 tested conversations. International expansion via Better Evidence Network.
📄 Claude lands in Microsoft Word
Anthropic's Office takeover completes. Excel and PowerPoint in February, Word now in public beta. Native sidebar, tracked changes, clickable citations that jump to the referenced section. Team and Enterprise subscribers first. Copilot is no longer the only adult in the building.
🏢 Microsoft Copilot goes agentic in Word, Excel, PowerPoint
Generally available. Multi-step actions, app-native. Word drafts, rewrites, restructures. Excel explores data, builds formulas, tables, visuals. PowerPoint creates and updates decks respecting templates. Sidebar shows every step. Microsoft reports +50% tries per user per week on Word and Excel. Copilot catches up to Claude on the same day Claude lands in Word.
📱 X ships Grok-powered Custom Timelines
Pin up to 10 niche feeds to your home tab. 75+ topics. Grok reads every post, tags it, layers on personalization. Premium iOS only at launch. Android "in the works." Feeds for cricket. Feeds for Formula 1. Feeds for ice hockey. No feed for "stop doomscrolling."
🎬 HeyGen HyperFrames turns HTML into MP4 for AI agents
Open source. Write HTML and CSS, render MP4 locally or in Docker. Headless Puppeteer + FFmpeg. Supports GSAP, Lottie, CSS transitions, Three.js, WebGL shaders. Deterministic output — same HTML, same video, every time. Installs as a Claude Code skill via `npx skills add heygen-com/hyperframes`. Slash commands: `/hyperframes`, `/gsap`. Agents can now ship motion graphics end to end.
🖼️ Ideogram ships Custom Branded Models
Train on 15 to 100 images. Quality over quantity — 40 curated captions beats a larger mediocre set. Products, characters, mascots, illustration styles, brand aesthetics. Pro and Team plans, API included. Enterprise for anything bigger. Your brand's visual identity, generation-ready.
🔓 Anthropic's "too dangerous" Mythos got leaked on day one
Anthropic announced Mythos Preview on April 7 — general-purpose frontier model, especially good at finding software vulnerabilities. Limited release: 40 handpicked companies including Microsoft, Apple, Google. Same day: a private Discord group guessed where it was hosted via a third-party vendor environment. Access achieved. One member: an Anthropic contractor. The group has been using it regularly. No evidence of cyberattacks yet. Altman's response on the Core Memory podcast: "It is clearly incredible marketing to say, we have built a bomb… we will sell you a bomb shelter for $100 million."
🤖 A Chinese robot just ran a half-marathon faster than any human ever has
Beijing E-Town Half-Marathon, April 19. Honor's "Lightning" humanoid: 50 minutes, 26 seconds. Seven minutes faster than the human world record. Four robots under an hour. Honor took first, second, third. Last year's winner finished in 2h40m. 100+ teams competed, up 5x from the debut year. Meanwhile in America: we held a marathon and Optimus showed up.
