AI news without the noise
Fast briefings on the models, agents, tools, and policy moves shaping AI in 2026 — sent when something important is trending, written for builders, founders, and curious people who want signal over hype.
Email alerts when AI news matters · Telegram available · unsubscribe anytime
Latest
Twenty-two pieces, no fluff. Sorted newest-first so you always hit what just dropped.
DeepSeek raised $7.4B at a $50B valuation. Qwen 3.7-Max scored 80.4% on SWE-Bench and runs autonomously for 35 hours. Chinese models now match US labs on capability while charging 4-29x less. But Washington is tightening the chip chokehold — and the real war has moved to silicon.
Read article →Anthropic's Fable 5 and Mythos 5 are the same model — what changes is the guardrails, not the intelligence. $10/$50 per million tokens, 1M context, SOTA vision and coding. Here's when to reach for it and when Opus 4.8 is still the smart pick.
Read article →Two Chinese frontier models now permanently charge $0.87 per million output tokens. DeepSeek scored 80.6% on SWE-bench Verified — 0.2 points behind Claude Opus. MiMo handles 1,000+ tool calls autonomously. Here's the real tradeoff for builders.
Read article →Anthropic's Mythos is now restricted to governments only — but hackers already breached it. OpenAI's AI solved an 80-year math problem. And a self-replicating worm poisoned 160 AI supply chain packages.
Read the briefing →Claude Code leads on SWE-Bench (69.2% vs 58.6%) and enterprise trust. Codex is faster, cheaper, and runs tasks in parallel. Benchmarks, pricing, real user feedback, and which one you should actually use.
Read article →Anthropic's latest flagship beats GPT-5.5 on 4 of 5 major benchmarks — agentic coding, reasoning, computer use, knowledge work. But it takes 30% more turns to win. And Mythos-class models are coming in weeks.
Read article →Anthropic restricted the most capable cybersecurity model ever built to governments and utilities. Here's the problem: an unauthorized group accessed it through a vendor breach in April. The restricted-AI era is here, and it's already broken.
Read article →OpenAI launched ChatGPT's bank-account feature two days after a federal class action accused the company of secretly sharing user data with Meta and Google. The feature is well-designed. The trust gap is not.
Read article →One of OpenAI's 11 co-founders just picked its biggest rival. Andrej Karpathy's move to Anthropic is more than a hire — it's the verdict on where the best people believe the most important work is happening.
Read article →AI accounted for 26% of all US job cuts in April — making it the top reason for layoffs for the second straight month. Cisco posted record revenue and cut 4,000 the same day. The efficiency dividend is here.
Read article →Hermes Agent just became the most-used AI app on OpenRouter — 224 billion daily tokens, ahead of every chat app and agent on the platform. The architecture that won, and what it means.
Read article →Two frontier models independently completed a 32-step corporate network intrusion. The offensive-defensive balance just shifted — and the defenders might actually be winning.
Read article →Rate limits doubled, peak-hour caps gone, and a SpaceX deal that puts "multiple gigawatts of orbital AI compute" on the roadmap.
Read article →Plugin migrations, silent delivery failures, memory hiccups, Windows quirks, and the relentless work that turned user frustration into shipped fixes.
Read article →Anthropic's biggest vertical push yet — purpose-built agents for pitchbooks, KYC, close, and more. Plus M365 add-ins and Moody's MCP app.
Read article →Three frontier models within four points of each other. What actually separates them — and which one to pick for your workload.
Read article →The seven tools we actually open every week — for thinking, coding, research, meetings, and audio. With honest pros and cons.
Read article →What changed in twelve months, what is actually working, where teams are still getting burned, and the architecture that wins.
Read article →Cursor for flow, Claude Code for the hard stuff, Windsurf for value. The two-tool workflow most senior engineers settle on.
Read article →Four frontier-class open-weights models in twelve days, at a third of the cost. Why self-hosting is no longer a quality downgrade.
Read article →Penalties up to €15M or 3% of global turnover. The 90-day compliance checklist for high-risk AI systems.
Read article →A clear explainer of what agents are in 2026, what they need to work, the three flavors people confuse, and where to start.
Read article →Why readers stay
"The only AI update feed that respects my time. Five minutes, no fluff, and at least one thing I act on."
"I forwarded the coding-agent comparison to my whole team. We changed our workflow the same day."
"The signal-to-hype ratio is unmatched. When AI Radar sends something, I know it is worth opening."
Topics
Six beats, one promise: every piece earns its read time.
Chat subscriptions
Prefer chat over email? Subscribe in Telegram and we’ll send AI Radar whenever something important is trending.
No email required. Tap the button, open Telegram, and press Start. The bot saves your chat and sends AI Radar updates there.
/stop in Telegram anytime to unsubscribe.
AI news without the noise. Signal only. Free.