DeepSeek V4 Drops as the Cheapest Frontier Model — 1/20th the Cost of Opus 4.7

🧠 LAUNCH

DeepSeek V4 Drops as the Cheapest Frontier Model — 1/20th the Cost of Opus 4.7

DeepSeek V4 lands with benchmark numbers rivaling Opus 4.7 and GPT-5.5 — at a fraction of the price. Million-token context built for agentic use, and the cost-performance ratio isn't incremental, it's a category disruption. If you're running inference-heavy workloads, your current provider bill just became indefensible without a re-eval. This is the most aggressive price-performance move since DeepSeek-R1 cratered the "scaling costs scale linearly" assumption. (4,047 likes | 309 RTs) Read more →

GPT-5.5 Hits the API With the Most Expensive Frontier Pricing Yet

GPT-5.5 and GPT-5.5 Pro are now available to developers — and the Pro tier at $30/1M output tokens is the priciest frontier API on the market. OpenAI is betting that raw capability justifies premium pricing even as DeepSeek undercuts everyone. The strategic contrast with DeepSeek V4 couldn't be sharper: one bets on volume, the other on margin. (4,189 likes | 214 RTs) Read more →

DeepSeek-V4-Pro open weights land on HuggingFace. The Pro variant — benchmarking at Opus 4.7 level — is downloadable today. The open-source frontier just leapfrogged again, and anyone with the hardware can now run a frontier-class model without API dependencies. (2,415 likes | 30 downloads) Read more →

OpenAI ships ChatGPT for Clinicians — free and domain-specific. OpenAI's second vertical healthcare product is free for clinicians, signaling this is a distribution play, not a revenue one. Get doctors hooked on Claude-competitor workflows before hospital procurement cycles kick in. (4,620 likes | 534 RTs) Read more →

Google dumps a week of Cloud Next AI infrastructure at once. Eighth-gen TPUs (TPUt/TPUi), Gemini Enterprise Agent Platform, Gemini Embedding 2 GA, agentic data cloud, Workspace Intelligence — the most comprehensive AI infrastructure update Google has ever shipped in a single week. If you're on GCP, there's a lot to absorb. (100 likes | 16 RTs) Read more →

💡 INSIGHT

Google in Talks to Pour Up to $40B Into Anthropic

Bloomberg reports Google is in discussions for up to $40 billion in Anthropic — which would be the largest AI investment in history. This would fundamentally reshape the competitive triangle between Google, Amazon, and Anthropic. Amazon already has a multi-billion stake; a Google deal of this size would make Anthropic the most heavily backed AI company on earth and raise serious questions about independence. Watch for deal confirmation and what it means for Anthropic's cloud-neutrality story. (250 likes | 313 RTs) Read more →

Anthropic and Amazon lock in 5 gigawatts of new compute. Five gigawatts is datacenter-city-level power — enough to run multiple hyperscale training clusters simultaneously. This isn't just about training the next model; it's about having enough inference capacity for an enterprise customer base that's scaling faster than anyone expected. Read more →

Meta goes to AWS for tens of millions of Graviton cores. Even the most vertically-integrated AI company on earth needs external compute. Meta tapping AWS Graviton for AI infrastructure signals that agentic workloads at billions-user scale exceed what any single company can build in-house. (1,018 likes | 79 RTs) Read more →

Musk vs. OpenAI heads to jury selection Monday. The AI era's biggest legal battle arrives in an Oakland courtroom next week. The outcome could force OpenAI's restructuring, threaten its $300B valuation, and set precedent for how AI companies can change corporate structure mid-flight. Whatever your take on Musk, the legal questions are genuinely novel. Read more →

Anthropic and NEC team up for Japan's largest AI engineering workforce. Anthropic's first major Japanese enterprise partnership — NEC is Japan's largest IT services company, and this signals a serious push into Asia's biggest enterprise AI market. Expect Claude enterprise pricing tailored for Japanese verticals. Read more →

🔬 RESEARCH

Anthropic's Project Deal: Claude Negotiates Real Transactions in an Office Marketplace

Anthropic built a real marketplace inside their San Francisco office and let Claude handle buying, selling, and negotiating with actual money on the line. This is the first controlled study of LLM economic behavior with real stakes — not simulated benchmarks, not toy games. The results reveal how Claude actually behaves when incentives are real: where it cooperates, where it pushes for advantage, and where its negotiation strategies diverge from human expectations. This is foundational work for anyone building agentic commerce. (3,772 likes | 251 RTs) Read more →

OpenAI launches a biosecurity-specific bug bounty for GPT-5.5. A dedicated biosecurity bounty for a single model release is unprecedented — OpenAI is treating GPT-5.5's capabilities as requiring domain-specific red-teaming beyond standard security review. If you have biosecurity expertise, this is worth your time. (1,719 likes | 124 RTs) Read more →

DeepMind's Decoupled DiLoCo trains across flaky data centers. The method tolerates network failures and cross-datacenter latency without synchronization bottlenecks — the key enabler for training models bigger than any single cluster can handle. As training runs stretch across continents, this becomes infrastructure-critical. (956 likes | 129 RTs) Read more →

🔧 TOOL

GPT-5.5 tops CursorBench at 72.8% — Cursor ships same-day integration. Cursor's own benchmark crowns GPT-5.5 as the best coding model, and they shipped integration the same day the API launched. The speed of ecosystem adoption tells you how seriously tool makers take this release — if you're a Cursor user, the switch is one dropdown away. (3,508 likes | 163 RTs) Read more →

Claude Code gets a web and mobile refresh with desktop file browser. The web and mobile interfaces get a significant UX overhaul — the standout is the desktop file browser with CMD+Shift+F for fast project navigation. The rapid iteration continues as Claude Code pushes into multi-surface territory. (1,752 likes | 76 RTs) Read more → For a deeper dive on getting the most out of Claude Code's growing feature set, see our guide on how to effectively prompt Claude Code.

Anthropic ships a Rate Limits API for programmatic org-level queries. Admins can now query rate limits per org and workspace via API — essential for teams managing API budgets at scale and building internal usage dashboards. Small feature, big operational unlock. Read more →

Sakana AI launches Fugu: multi-agent orchestration in beta. Sakana's first commercial product — a multi-agent orchestration system from the lab known for evolutionary model merging. Fugu sits between single-agent tools and full platform plays, offering a lighter-weight entry point for teams not ready for Managed Agents. (478 likes | 121 RTs) Read more →

📝 TECHNIQUE

Claude Code setup plugin auto-configures hooks, skills, and MCP for your project. An Anthropic plugin that analyzes your codebase and recommends which Claude Code automations to activate — hooks, skills, MCP servers, subagents. Dramatically lowers the setup barrier for teams adopting Claude Code's full feature set instead of just the chat interface. (1,445 likes | 141 RTs) Read more →

Qwen3.6-27B runs on a Raspberry Pi coding a web app live. A frontier-class 27B model running on a Pi, coding a working app in real-time. The gap between local AI and cloud AI capabilities keeps collapsing at the edge — and this demo makes the point viscerally. (3,440 likes | 283 RTs) Read more →

🏗️ BUILD

Superpowers: the 166K-star agentic skills framework. The most popular new agent framework this month codifies the "skills" pattern — reusable, composable agent capabilities — into a methodology and library. If you're building agentic systems and tired of reinventing skill definitions, browse the library before writing your own. (166,764 likes | 14,657 RTs) Read more →

🎓 MODEL LITERACY

Mixture of Experts (MoE) and Inference Cost: DeepSeek V4 achieves frontier performance at 1/20th the cost of Opus 4.7 — and the secret is Mixture of Experts architecture. Instead of running every token through the entire model, MoE routes each token to a small subset of specialized "expert" sub-networks. A 600B-parameter MoE model might only activate 30B parameters per token, slashing compute costs while maintaining quality. This is why cost-performance disruptions keep coming from architecture innovation, not just scaling up. When you see a model that's "cheap but good," check whether it's MoE — the answer is almost always yes.

⚡ QUICK LINKS

Transformers.js in Chrome Extensions: Step-by-step guide to running ML models in browser extensions — no server, no API keys. Link
Browser Harness: Open-source browser automation designed for LLM agents — lightweight alternative to Playwright-based approaches. (73 likes | 28 RTs) Link
Anthropic TypeScript SDK v0.91.1: Security patch for CMA memory file permissions — update immediately if using managed agent memory in production. Link

🎯 PICK OF THE DAY

Project Deal reveals that the real test of AI agents isn't benchmarks — it's economic behavior under real stakes. Anthropic built a marketplace inside their office where Claude handles actual buying, selling, and negotiating with real money. This isn't another eval suite or simulated environment — it's the first controlled study of how an LLM behaves when incentives are genuine. And the results matter far beyond research novelty. As we barrel toward agentic commerce — AI agents booking travel, negotiating SaaS contracts, bidding in auctions — the gap between how LLMs negotiate and how humans expect them to may be the biggest unsolved alignment problem for commercial AI. Do you want your purchasing agent to maximize savings aggressively, or maintain relationships? To accept a fair price quickly, or counteroffer three times? These aren't technical questions — they're economic and social ones, and Project Deal is the first serious attempt to map that terrain with real data. If you're building anything where an AI agent touches money, this paper is required reading. (3,772 likes | 251 RTs) Read more →

Until next time ✌️