Tokens & Signals

Tokens & Signals for 6/23/2026. We scanned ~1,200 Twitter accounts (1201 tweets), 13 subreddits (51 posts), Hacker News (13 stories), 5 newsletter posts, 3 podcast episodes, 136 Discord messages, and leaderboard data for you. Estimated reading time saved: ~10 hours.

TLDR & AI Twitter Recap

Anthropic launched Claude Tag for Slack—it's now a persistent team member with its own identity, memory, and access to your codebase. 65% of Anthropic's own code is already generated this way. x.com/bcherny/status/2069474681749754272

DeepSeek raised $7.4B at a $60B valuation, with CEO Liang Wenfeng personally throwing in $3B to lock in control and keep talent from walking. x.com/teortaxesTex/status/2069336274532479049

The Five Eyes alliance is warning that frontier AI capable of serious cyberattacks is months away, not years. x.com/kimmonismus/status/2069446992200187912

Mistral dropped OCR 4, which is crushing benchmarks at 85.2% and outputting structured data like bounding boxes for RAG pipelines. x.com/MistralDevs/status/2069421708142952761

Krea 2 is live: a 12B parameter open-weights image model with a 2-second "Turbo" mode. huggingface.co/krea/Krea-2-Turbo

@sharifshameem on VibeThinker: "A 3B model beating Opus 4.5 on reasoning? The efficiency gains are getting wild." x.com/sharifshameem/status/2069487039909925121

@pmddomingos on the "science slop" trend: "When you rely on AI to write your code, you risk losing the foundational understanding of the systems you're building." x.com/pmddomingos/status/2069170489684373837

OpenAI DevDay 2026 is September 29 in SF. Applications close July 10. x.com/OpenAI/status/2069483224158646739

Google invested $75M in A24 to develop custom AI filmmaking tools. reddit.com/r/artificial/comments/1ud44pc/google...

CXMT (Chinese DRAM maker) is closing the price gap with Samsung and SK Hynix fast — they're not playing the budget game anymore. x.com/jukan05/status/2069488184036307049

Best to Build With Today

Coding — claude-opus-4-8-thinking (Current LiveBench leader).

Reasoning — gemini-3.1-pro for hard logic and math tasks.

Chat — gemini-3.1-pro for general assistant work.

Image generation — Krea 2 Turbo for speed and high-quality local generation.

Embeddings — Mistral OCR 4 for structured document ingestion and RAG optimization.

Deeper Dives

💼 Industry & Business

DeepSeek's $7.4B Funding Round

DeepSeek closed its first external round at $7.4B and a $60B valuation. CEO Liang Wenfeng personally put in roughly half the round to keep control tight and give talent something to stick around for.

* Why it matters: It's a serious signal that DeepSeek is swinging for AGI dominance — and betting efficient model distillation is the way to get there.

� Twitter� Reddit

CXMT Challenging DRAM Leaders

Chinese chipmaker CXMT is now pricing its DRAM only 5–10% below Samsung and SK Hynix. With 700% YoY revenue growth in Q1 2026, they've stopped being a cheap alternative and started being an actual competitor.

* Why it matters: They're not undercutting anymore — they're matching. That's a very different threat.

� Twitter� Newsletter

Google Invests $75M in A24

Google is putting $75M into A24 to build custom AI tools specifically for filmmakers and creators.

* Why it matters: Generative video tech is about to land in the hands of people who actually know how to make great films. That's a big deal.

� Reddit

🧠 Models & Research

Five Eyes AI Cyber Warning

Intelligence agencies from the US, UK, Canada, Australia, and New Zealand are sounding the alarm: frontier AI will fundamentally reshape offensive and defensive cyber capabilities — and it's coming within months, not years.

* Why it matters: AI cyber risk just graduated from "tech industry concern" to full-blown national security issue.

� Twitter� Reddit

VibeThinker 3B

VibeThinker is a 3B parameter model that claims to beat much larger frontier models on reasoning benchmarks, using SFT+GRPO training to punch way above its weight class.

* Why it matters: Serious reasoning capability is getting compressed into tiny, cheap models. The big players should be paying attention.

� Hacker News� Reddit

🚀 Products & Launches

Anthropic 'Claude Tag'

Anthropic's new Slack integration lets Claude act as a persistent teammate — it remembers project history, handles tasks on its own, and actually lives inside your workflow.

* Why it matters: This isn't a chatbot you open in a tab anymore. It's an agentic collaborator that shows up where you already work.

� Twitter� Hacker News� Reddit

Mistral OCR 4

Mistral's latest model is built for structured data extraction — bounding boxes, table classification, the works — to make document ingestion actually reliable.

* Why it matters: RAG pipelines are only as good as the documents feeding them. This is the tool that makes business docs readable.

� Twitter� Hacker News

Funding & Deals

DeepSeek raised $7.4B to accelerate AGI development and talent retention.

Google invested $75M into A24 for AI-powered filmmaking tools.

Launches

Claude Tag — An active, memory-enabled AI teammate for Slack.

Krea 2 — A high-performance 12B open-weights image model with a 2-second 'Turbo' variant.

Mistral OCR 4 — A document-processing model specialized for structured data extraction.

Closing thought: Between Anthropic turning AI into a Slack coworker and intelligence agencies ringing the alarm on cyberattacks, the biggest AI battles have quietly shifted away from chatbots. The new frontlines are infrastructure and national security.

DeepSeek’s $60B Bet on AGI Dominance