Anthropic ships Claude Tag, an async Opus 4.8 teammate for Slack

Good morning 👋 Anthropic just put Opus 4.8 into your Slack and told it to keep working while you're in other meetings — for days if needed.

In today's issue:

• 🔭 Anthropic ships Claude Tag: async Opus 4.8 in Slack
• 🧠 Mistral's OCR 4 covers 170 languages, $2/1K on batch
• 🔬 Localization is the hidden tax eating half your code agent's token budget
• 🚀 Superhuman acquires GPTZero; Engram launches with $98M
• 🔗 Agent billing, DeepMind x A24, and a three-year T-cell mystery solved

🔭 THE ONE THING

🏷️ Anthropic ships Claude Tag, an async Opus 4.8 teammate for Slack

Anthropic launched Claude Tag today, an Opus 4.8 Slack integration built for async delegation: @-tag it, it stages the work, reads channel history, chains tools, and reports back. Beta for Enterprise and Team customers, Slack-only, expansion to other platforms planned but undated. Their lead stat: "65% of our product team's code is created by our internal version of Claude Tag." Generated, not shipped, and they're the seller, so treat the magnitude as directional rather than precise. A live internal production figure is still a different category than a curated benchmark, and the underlying UX bet is real: "assign" lands differently than "ask," which is what makes delegation actually happen.

🧠 MODELS & RELEASES

• 📄 Mistral AI shipped OCR 4, covering 170 languages with per-word confidence scores, bounding boxes, and block classification across content types (titles, tables, equations, signatures). OmniDocBench 93.07. API is $4/1K pages or $2 on batch, and it runs in a single container if data sovereignty means you have to self-host.

🔬 RESEARCH HIGHLIGHTS

• SHERLOC — LLM coding agents spend roughly half their token budget just locating faults before writing a single edit. SHERLOC replaces naive file-retrieval localization with structured diagnostic output: it returns the fault site plus a diagnosis of *why* that location is the problem. That context is what a repair agent actually needs. If you're building or benchmarking code agents, localization quality is the hidden variable most evals ignore. arxiv

• InSight — VLA robots are capped by their training demos: if a task wasn't in the dataset, the robot can't do it. InSight makes VLAs steerable at the primitive-action level, letting them chain novel behaviors from text instructions without new demonstrations. Autonomous skill acquisition from language alone, in a physical robot, is a meaningful step past imitation. arxiv

• OpenThoughts-Agent — Most open agentic training efforts (SWE-Smith, SERA, Nemotron-Terminal) optimize for a single benchmark. This paper releases data recipes for training broadly capable agents across task types. The open release is the point. Leaderboard-specific training is why so many "agentic" models fall apart the moment you give them a real task. arxiv

🚀 AI STARTUPS

• Superhuman is acquiring GPTZero, Edward Tian's AI-detection platform. Terms undisclosed. The fit on paper: Superhuman helps you write with AI; GPTZero flags what AI wrote. They're calling the combined thing an "authenticity layer" embedded in Superhuman Go. Whether that tension is a clever bundle or a contradiction depends on how cynically you read enterprise software roadmaps.

• Engram launched with $98M from General Catalyst, Kleiner Perkins, and Sequoia to solve enterprise context at the model level. The premise: stop making every agent call re-read the same company docs from scratch. Engram compresses org knowledge into compact, reusable model memory and claims 1–10% of the usual token spend. Andrej Karpathy advising; Chris Ré (Stanford) co-founding; Microsoft 365, Notion, and Harvey already testing it. Good team, real problem. Valuation undisclosed.

🔗 QUICK LINKS

• Ampersend built a pay-per-intelligence routing layer on Bedrock AgentCore where agents pick models, bill per request, and handle two-hop agent-to-agent payments automatically. link

• Google DeepMind announced a research partnership with A24 to build AI creative tools with filmmakers at the table while the technology is still taking shape. link

• OpenAI published immunologist Derya Unutmaz's account of using GPT-5 Pro to crack a T-cell question that had stalled his lab for three years. link

That's what mattered today.

Pradeep Perugu