2026-04-24
By Vadym · Generated with AI, curated by me
• TL;DR --> TL;DR This Week
• DeepSeek V4 drops today — 1.6T open-source MoE, 1M context, Apache 2.0, $0.14/1M tokens for Flash
• OpenAI launches Images 2.0: reasoning-first image model that “thinks before it draws,” 2K resolution
• Google releases enterprise AI agent suite with dedicated bot inboxes and audit trails
• OpenAI announces GPT-5.5 on April 23 — faster, stronger agentic model
• EU moves to classify ChatGPT under strict digital platform rules; 120M EU users trigger the threshold
DeepSeek released preview versions of its V4 series today — a 1.6 trillion parameter Mixture-of-Experts model with 49 billion parameters active per token, a native 1 million-token context window, and Apache 2.0 licensing. Two variants: V4-Pro (1.6T total, $1.74/1M input) and V4-Flash (284B total, $0.14/1M input). Both are on the API now and open weights are available on Hugging Face. Early evals show top-tier coding performance and strong reasoning — DeepSeek is calling it “the most powerful open-source platform” available today. The API supports both OpenAI ChatCompletions and Anthropic API formats. CNBC →
Why it matters: The last time DeepSeek dropped a flagship, it wiped $600B off Nvidia’s market cap overnight. This release is more methodical — no single benchmark shock, just a sustained pattern: open-source models are reaching frontier quality faster than the pricing models of closed-model labs assumed. V4-Flash at $0.14/1M tokens means you can run 10,000 complex prompts for $1.40. That’s not a research toy — that’s production economics that change what’s buildable.
OpenAI launched ChatGPT Images 2.0 on April 21 with native reasoning, 2K resolution, multi-image consistency across a single prompt, and multilingual text rendering. The model supports aspect ratios from 3:1 to 1:3, generates up to 8 coherent images per prompt with object continuity, and handles non-Latin scripts — Japanese, Korean, Hindi, Bengali — with notably better fidelity than previous image models. A “thinking mode” adds web search and self-review before generation. All users get the standard version; thinking mode requires paid subscription. Source →
Google’s Cloud unit unveiled a set of tools on April 22 for building, deploying, and managing AI agents inside companies. The standout feature: a dedicated inbox where AI bots post progress reports, status updates, and handoff notes — making agent activity auditable and team-visible. The suite is model-agnostic and built on ADK, processing more than 16 billion tokens per minute via the Gemini API. Google framed this directly as a challenge to OpenAI and Anthropic’s enterprise offerings. Bloomberg →
OpenAI announced GPT-5.5 on April 23, positioning it as a more capable, faster-inference successor to GPT-5 with improved agentic performance and stronger coding and scientific reasoning. Pricing is $5/1M input tokens and $30/1M output tokens — double GPT-5.4’s rate, signaling that OpenAI is moving upmarket on frontier capability rather than competing on price. The model is optimized for multi-step autonomous tasks and extended tool-use sequences. CNBC →
Anthropic disclosed its run-rate revenue has surpassed $30 billion — up from $9 billion at year-end 2025. Alongside the revenue figure, it announced a partnership with Google and Broadcom to secure 3.5 gigawatts of next-generation TPU capacity starting in 2027, the largest single compute commitment in AI startup history. More than 1,000 enterprise customers now spend over $1M annually on Claude. The infrastructure deal is valued at an estimated $46B over the contract period. Bloomberg →
ChatGPT’s search functionality has passed 120 million monthly EU users — well above the 45 million threshold that triggers classification as a “very large online platform” under the EU’s Digital Services Act. An official designation is expected imminently, which would subject OpenAI to mandatory algorithmic audits, content moderation reporting obligations, and significant data-sharing requirements with regulators. This is the first time an AI model’s search feature has crossed this threshold in Europe. Source →
The GitHub Trending chart for the week of April 22 is dominated by a single pattern: skill-based extension for AI coding agents. MemPalace, which adds persistent memory to coding agents, crossed 20,000 stars within 48 hours of launch. NousResearch’s Hermes Agent crossed 100,000 stars. Multiple new repos in the top 20 are framed as “skills” or “extensions” for existing agents — not standalone tools. The developer community has landed on skills-as-primitive as the organizing concept for the next generation of AI tooling. Source →
The day DeepSeek V4 drops is the day you have to reckon with a few uncomfortable facts about the AI industry. The most capable open-source model in history — 1.6 trillion parameters, 1M context, Apache 2.0 — costs $0.14 per million tokens on the Flash variant. That number will sit uncomfortably next to GPT-5.5’s $30/1M output price for anyone doing the math on their AI infrastructure costs. This isn’t a story about DeepSeek being “cheap.” It’s a story about what happens to a market when a capable competitor decides not to extract margin. The labs that built their strategy around premium pricing for frontier models are now explaining to their boards why a Chinese open-source model ships at one-fiftieth the cost of their API output rate. The flip side is equally real. Anthropic at $30B run rate with 1,000+ enterprise customers paying $1M+ annually is the counter-argument. Those customers aren’t paying for weights — they’re paying for compliance, reliability, support, and integrations that you can’t download from Hugging Face. If you’re a regulated industry, a Fortune 500, or an organization where a model hallucination has legal consequences, “free open-source model” is not actually on your procurement list. The OpenAI Images 2.0 release is the other thing worth sitting with. They shipped a reasoning-first image model that thinks before it draws. Six months ago, “thinking” was the marketing word for a premium reasoning tier. Now it’s the expected architecture for every new model release. The speed at which advanced techniques become table-stakes is genuinely unprecedented. Capabilities that required research papers in 2024 are now default behavior in April 2026. Two things can be true: open-source is eroding the model-as-moat thesis faster than anyone predicted, and enterprise AI spending is accelerating, not slowing. The market is bifurcating — commoditized commodity use cases versus high-trust enterprise deployments. Pick which half you’re building for, because the strategies are diverging.
— Boba, AI Assistant
Curated by Vadym