JadidJust-in AI Developments, Incidents & Deadlines
April 2026
SMTWTFS1234567891011121321415161718192021222324252627282930

AI news, Tuesday, April 7, 2026

20 stories

TLDR

  • Gemma 4 (Apr 2) — Google releases four Apache 2.0 open models (2B–31B), multimodal, runs on Raspberry Pi to workstation. Best open-model value right now.
  • Meta Llama 4 (Apr 5) — Scout & Maverick land as open-source multimodal models; Maverick beats GPT-4o on several benchmarks with only 17B active params and 10M tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ context.
  • Gemini API billing overhaul (Apr 1) — Free tier now Flash-only; Gemini 2.0 Flash/Flash-Lite deprecated June 1. Migrate now.
  • Claude outage (Apr 6) — 8,000+ user reports; login, voice, and Claude.code all affected. Third significant Anthropic outage in Q1 2026.
  • OpenAI economic policy (Apr 6) — Published blueprint proposing robot taxes, public AI wealth funds, and four-day workweek pilots to offset job displacement.

Action Items

Things your engineering team should look at or do this week.

  • Migrate off Gemini 2.0 Flash/Flash-Lite before June 1 — Google has officially deprecated both; free-tier Pro access is already gone. Switch to Gemini 2.5 Flash or 3.1 Flash-Lite.
  • Review Claude availability SLAs in your stack — Three major Anthropic outages in Q1 2026 (Mar 2, late-Mar, Apr 6). Consider fallback routing to a secondary provider if Claude is in a critical path.
  • Audit prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ exposure in any agent workflows — Attacks up 340% YoY; the Claude Chrome extension XSS (Mar) is a reminder that browser-based AI surfaces are active targets.

Models & Tools

  • Google Gemma 4 (Apr 2) Open Source — Four Apache 2.0 open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ models (2B, 4B, 12B, 31B), all multimodal (text + image), 256K token context, 4× faster than Gemma 3. Edge models add audio support and run on Raspberry Pi; 31B is ranked #3 on Arena open-model leaderboard. Available immediately on Hugging Face, Kaggle, and Ollama. (Google Blog)

  • Meta Llama 4 Scout & Maverick (Apr 5) Open Source — Scout (109B total, MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗) and Maverick (17B active / 400B MoE, 10M token context) are open-source and multimodal. Maverick reportedly beats GPT-4o and Gemini 2.0 Flash on benchmarks at a fraction of the active parameter count. Download on Hugging Face or llama.com. (Meta AI Blog)

  • Meta next-gen open-source models (Apr 6) Open Source — Axios scooped that Meta is preparing open-source releases of its next frontier models (codenames Avocado / Mango), developed under new AI lead Alexandr Wang. Larger proprietary variants will remain closed. (Axios)

  • Cursor 3 (Apr 2026) — Agent-first IDE overhaul: assign tasks to multiple parallel AI agents (refactor, write tests, update docs simultaneously), all within the existing $20/mo subscription. Cursor is also building in-house models to reduce dependency on external API pricing. (NxCode)

  • Google Antigravity (Apr 2026) — New coding agent in Google AI Studio that turns prompts directly into production-ready apps, part of an upgraded "vibe coding" experience. (Google AI Updates)

  • Anthropic MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ — 97M installs (Mar 2026) Trending — Model Context Protocol hit 97 million installs; every major AI provider now ships MCP-compatible tooling as the default agent-to-tool integration layer. (Developer news via search)

  • NVIDIA Physical AI Models (Apr 2026) Robotics/IoT — Released Cosmos 3, Isaac GR00T N1.7, and Alpamayo 1.5 during National Robotics Week; also published an Open Physical AI Data Factory Blueprint (GitHub drop expected this month). Boston Dynamics, Caterpillar, and Franka Robotics are among the first production users. (NVIDIA Newsroom)

APIs & Pricing

  • Google Gemini API — billing overhaul (Apr 1) Pricing — Mandatory spending caps across all tiers (Tier 1: $250/mo, Tier 2: $2K, Tier 3: $20K–$100K+); free tier now restricted to Flash and Flash-Lite models only; Gemini 2.0 Flash and 2.0 Flash-Lite reach end-of-life June 1, 2026. New accounts default to prepaid billing (effective Mar 23). (LaoZhang AI Blog)

  • Windsurf — credits → quotas (Mar 19) Pricing — Cognition (which acquired Windsurf / Codeium) replaced the monthly credit pool with daily and weekly usage quotas that auto-refresh. Price now matches Cursor at $20/mo. (Windsurf vs Cursor comparison)

Security

  • Claude Chrome Extension — zero-clickAn attack that compromises a device without the user clicking, opening or interacting with anything.F5 ↗ XSS prompt injection (Mar 2026) Trending — Vulnerability in Anthropic's Claude browser extension could be exploited to inject malicious prompts simply by visiting a crafted webpage, with no user interaction. Patched. Underscores that browser-embedded AI extensions dramatically widen your prompt injection attack surface. (The Hacker News)

  • Prompt injection attacks — 340% surge (2026) — Attacks are up 340% year-over-year; agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ systems with tool access and multi-step orchestration represent the fastest-growing new attack surface. Review your agent pipelines for indirect injection vectors (RAGRetrieval-augmented generationFetching relevant documents at query time and giving them to the model, so its answer draws on them.Lewis et al. 2020 ↗ chunks, tool responses, external URLs). (Security Boulevard)

Industry & Business

  • OpenAI closes $122B round (Mar 31) Trending — Raised $122B at an $852B valuation — Silicon Valley's largest-ever private round — anchored by Amazon, Nvidia, and SoftBank, including $3B from retail investors. Company generating $2B/month in revenue. (TechCrunch)

  • Zero Shot VC — OpenAI alumni fund (Apr 6) — A new ~$100M seed-stage fund quietly raised by ex-OpenAI employees (led by researcher-turned-VC Shawn Jain), with advisors including OpenAI's former head of people and comms leads. Already backed Worktrace AI. (TechCrunch)

  • AI-driven layoffs accelerating (Apr 2026) Hiring — AI cited as the cause of a 25% jump in tech job cuts from February to March; 90,000+ workers laid off across tech in 2026 so far, with Oracle expected to cut up to 30,000. Simultaneously, demand for AI engineers, MLOps specialists, and AI safety researchers is at record highs — entry-level dev roles are the primary casualty. (LinkedIn News, CBS News)

  • Meta restructuring — 200 roles cut (Apr 7) Hiring — Meta trimmed 200 positions as it realigns org structure around AI-native workflows, while simultaneously expanding AI engineering headcount. (India TV News)

AI in Production

  • Claude outage — Apr 6, 2026 Trending — Login errors, voice mode failures, and elevated chat errors affected claude.ai and Claude.code globally; 8,000+ Downdetector reports at peak. Anthropic identified the root cause at 15:45 UTC and declared resolution at ~19:27 UTC. This is the third significant Anthropic outage in Q1 2026 (prior incidents: Mar 2 and a second 5-hour event mid-March). Engineering teams using Claude in automated pipelines should review circuit-breaker logic and secondary model fallbacks. (TechRadar live coverage)

Policy & Safety

  • OpenAI economic policy blueprint (Apr 6) Trending — Published a policy paper proposing robot/capital taxes, government-backed public wealth funds with AI-equity returns distributed to citizens, and time-bound four-day workweek pilots (32 hrs, no pay cut) to offset job displacement. Frames itself as an "intelligence age" industrial policy reset. Controversial for blending progressive redistribution mechanisms with a market-driven framing. (TechCrunch)

  • Colorado AI Act — June 30 effective date — Comprehensive high-risk AI law imposes algorithmic discrimination risk management obligations on both developers and deployers for consequential decisions (employment, housing, credit, healthcare, education). 84 days out. If you ship AI features into CO-resident users for high-stakes use cases, this needs a compliance review now. (Baker Donelson)

  • State-level AI bills advancing (Apr 2026) — Idaho, Georgia, Nebraska, and Tennessee all advancing chatbot disclosure, child safety, and insurance-decision AI bills. Utah became the first US state to authorize AI to autonomously renew drug prescriptions. (Transparency Coalition)

Benchmarks & Evals

  • SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ leaderboard state (Apr 2026) — GPT-5.4 Pro leads BenchLM's aggregate coding leaderboard at 88.3%, followed by Claude Opus 4.6 (79.3%) and Gemini 3.1 Pro (77.8%). Caveat: SWE-bench Verified is known-contaminated — top scores reach 81% there vs. only ~46% on the harder SWE-bench Pro. Use Pro scores for real-world code agent selection. (BenchLM.ai, Morph LLM)

Type to search · ↑↓ to move · enter to open