OpenAI bets on biotech, Anthropic turns down $800B, and AI coding hits overdrive
Mon, Apr 20, 2026 · 10 stories
Anthropic turned down $800B in funding this week while the NSA reportedly used their Mythos model against policy. What do you do when your product is simultaneously "too dangerous to release" but also "too important to restrict"? But Mythos isn't the only signal that AI is going vertical. OpenAI launched GPT-Rosalind, their new life sciences model optimised for drug discovery and genomics. The next wave of AI isn't general-purpose chatbots; it's specialized models solving domain-specific problems.
The money agrees. Cursor raised $2B at a $50B valuation, Factory hit $1.5B, and OpenAI expanded Codex with desktop control and memory. AI coding is now the hottest single category in venture capital, and developer tooling history is being made as we speak.
Top Stories
Bloomberg Technology
Anthropic turned down multiple investor offers valuing the company at $800 billion or higher, while its Mythos model expanded to UK banks and US federal agencies. The NSA is reportedly using Mythos despite Anthropic's placement on a government procurement blacklist, and Dario Amodei met with White House Chief of Staff Susie Wiles to negotiate broader government access.
Why this matters:
Turning down $800B in an industry where every lab is raising aggressively signals Anthropic believes it can command an even higher number. Meanwhile, the Mythos government access story evolved from "too dangerous to release" to "too important not to deploy." The NSA usage despite blacklisting shows that when national security capabilities are at stake, procurement rules bend. For founders building in regulated verticals, the playbook is clear: capability-gated access is becoming the new go-to-market.
OpenAI Blog
OpenAI introduced GPT-Rosalind, a specialized frontier reasoning model designed to accelerate drug discovery, genomics analysis, and protein reasoning. It's OpenAI's first model purpose-built for a scientific vertical rather than general-purpose use.
Why this matters:
This is the clearest signal yet that the next wave of AI competition isn't about who has the best chatbot -- it's about who owns the vertical. Google has AlphaFold. Now OpenAI has Rosalind. Specialized models that speak the language of a domain (protein structures, genomic sequences, molecular interactions) will be harder to displace than general-purpose assistants. For AI founders, the lesson is that horizontal is table stakes; the defensible moat is vertical depth.
Bloomberg Technology / TechCrunch AI
Cursor is raising approximately $2 billion at a $50B+ valuation with a16z and Thrive Capital leading. The same week, Factory hit a $1.5B valuation for enterprise AI coding, OpenAI shipped a major Codex update with desktop control and memory, and VC dealmaking hit a record $267B in Q1 2026 with nearly all funds flowing to AI.
Why this matters:
The AI coding tool category went from "nice to have" to the single largest concentration of VC dollars in the industry. Cursor at $50B means investors are pricing in a world where AI writes the majority of production code. The simultaneous Codex expansion and Factory's enterprise play suggest we're about to see a Notion/Slack-style consolidation war in developer tooling -- and the winner gets the most valuable distribution channel in software.
Quick Hits
TechCrunch AI
Anthropic launched Claude Design, letting founders and PMs create quick visuals without design expertise.
Cloudflare Blog
Cloudflare launched an AI Platform with inference optimized for agent workloads, plus an email service enabling AI agents to send and receive emails programmatically.
Qwen Blog
Alibaba released two models: a 35B-parameter open-source agentic coding model and the Qwen3.6-Max-Preview flagship. Simon Willison reported the smaller model beat Claude Opus 4.7 at image generation tasks running locally on a laptop.
TechCrunch AI
OpenAI discontinued Sora and folded its science team, with Kevin Weil and Bill Peebles departing as the company pivots from consumer moonshots to enterprise AI.
claudecodecamp.com / tokens.billchambers.me
Independent analysis found Claude Opus 4.7's tokenizer inflates per-session costs by 20-45% compared to 4.6, the first rigorous community benchmark of model-to-model cost inflation.
The Verge
Vercel disclosed a security breach affecting internal systems, facilitated through a compromised third-party AI tool, with hackers claiming to sell stolen employee data.
The Verge
AI-driven RAM demand is expected to outstrip supply through 2027-2030, with Samsung, SK Hynix, and Micron only meeting 60% of demand by end-2027.
