AI Daily Digest — April 25, 2026 ⏱️ 8 min read
Today’s digest: OpenAI shipped GPT-5.5 — its first model designed as an agent runtime, not a chatbot — and DeepSeek fired back hours later with V4, the most capable open-source model yet. It’s a two-front war now.
🔥 Top Stories
OpenAI launches GPT-5.5: the “agent-first” model
Six weeks after GPT-5.4, OpenAI dropped GPT-5.5 — the first ground-up retrain since GPT-4.5. It tops the Artificial Analysis Intelligence Index at 60 and hits 82.7% on Terminal-Bench 2.0. The real shift: this isn’t a chat model with agent features bolted on. It’s an agent runtime that happens to chat. API pricing lands at $5/$30 per million tokens (input/output), with a Pro tier at $30/$180. Already live for Plus, Pro, Business, and Enterprise users. If you’re building agents, the Claude vs OpenAI calculus just shifted again. Read more →
DeepSeek drops V4: open-source is closing the gap fast
Hours after OpenAI’s announcement, DeepSeek released V4-Pro and V4-Flash. V4-Pro trails GPT-5.4 and Gemini 3.1-Pro by “3 to 6 months” according to independent benchmarks — which means it’s already better than what was state-of-the-art in late 2025. And it’s fully open-weight on Hugging Face. The agentic capabilities got a major upgrade too. For teams weighing open model APIs, V4-Flash looks like the new default. Read more →
Meta and Microsoft cut 20,000 jobs as AI reshapes headcount
Over 92,000 tech workers laid off in 2026 so far. Meta is cutting 10% of its workforce (~8,000 jobs) starting May 20, and Microsoft is trimming another 12,000. The companies are simultaneously pouring billions into AI infrastructure. The message is clear: fewer people, more compute. This isn’t a blip anymore — it’s a structural shift. Read more →
Musk’s Terafab taps Intel for AI chips
Tesla, SpaceX, and xAI are forming a joint venture to use Intel’s advanced chipmaking tech for the Terafab facility. Musk betting on Intel over TSMC is… a choice. Could be a smart geopolitical hedge, could be wishful thinking about Intel’s foundry roadmap. Either way, the AI chip supply chain just got more interesting. Read more →
Anthropic’s “too dangerous to release” Mythos model gets breached
A group gained unauthorized access to Anthropic’s Mythos model — the one deemed too powerful for public release — through a third-party vendor. No evidence of Anthropic’s own systems being compromised, but the irony is thick. Meanwhile, Anthropic’s Project Glasswing (an AI that finds software vulnerabilities) discovered bugs across every major OS and browser that survived decades of human audits. Make of that what you will. Read more →
🛠️ New Tools & Releases
| Tool | What’s New | Link |
|---|---|---|
| GPT-5.5 API | Live in API as of April 24. 1M context, $5/$30 pricing. First agent-native flagship. | openai.com |
| DeepSeek V4 | V4-Pro and V4-Flash on Hugging Face. Open weights, strong agentic reasoning. | euronews.com |
| OpenAI Codex | Now powered by GPT-5.5 on NVIDIA infrastructure. Agentic coding workflows. | nvidia.com |
| @wordpress/build | Replaces webpack+Babel with esbuild. Auto-generates PHP registration files. | wordpress.org |
| iOS 26.2 SDK | Alternative app marketplaces and external payment processing now available. | developer.apple.com |
💰 Funding & Business
- Q1 2026 broke every record: $300B into 6,000 startups globally. AI accounted for 80% ($242B). Four mega-rounds — OpenAI ($122B), Anthropic ($30B), xAI ($20B), Waymo ($16B) — absorbed 65% of all global VC.
- Cursor raising $2B at $50B+ valuation. For an AI code editor. Let that sink in.
- Eclipse disclosed $1.3B across two funds targeting AI infrastructure, manufacturing, and defense.
- Tesla earnings call: Terafab JV with SpaceX and xAI confirmed, using Intel foundry tech. Musk called it “the largest AI training facility ever conceived.”
📊 What Developers Are Discussing
- “AI Psychosis” is real. Karpathy says he hasn’t written a line of code since December. Spends 16 hours/day commanding agent swarms. Harvard research shows 14% of AI power users experience “brain fry” — 39% more serious errors, 33% more decision fatigue. The New Stack has the breakdown.
- GPT-5.5 as agent runtime vs. chat model — HN is split. Some say it’s the right abstraction. Others say OpenAI is chasing a market that doesn’t exist yet. Hot take: both are right.
- DeepSeek V4 open weights sparking another round of “is open-source AI actually catching up?” debates. The benchmarks say yes. The vibes say… maybe.
- LMDeploy critical vulnerability exploited within 13 hours of disclosure (CVSS 7.5). If you’re self-hosting LLMs, patch now. Don’t wait until Monday.
- 50+ state AI bills in play. California’s EO N-5-26 directing agencies to draft AI safety requirements. The regulatory wave is here whether we like it or not.
📝 Worth Reading
- OpenAI launches GPT-5.5 just weeks after GPT-5.4 — Fortune’s deep dive on the accelerating release cadence and what it means for the industry.
- AI agents are scrambling power users’ brains — Axios on the addiction loop: “They operate like slot machines.” Worth reading if you use Claude Code or Cursor daily.
- Q1 2026 shatters VC funding records — $300B quarter. The numbers are genuinely hard to process.
- AI Legislative Update: April 24 — 50+ state bills tracked. If you ship AI products in the US, this is required reading.
- From Code to Conductor: the Karpathy inflection — Thoughtful essay on developers evolving from writers to orchestrators of AI agents.