AI Morning Briefing — February 7th, 2026

Your daily digest of what's happening in AI, straight from the trenches.

---

🚀 Headlines (30 sec read)

• Claude Opus 4.6 beats Gemini 3 Pro — Anthropic's latest model scores 14 points higher in Text Arena
• Goldman Sachs deploys Claude AI — Automating accounting and compliance work with Anthropic engineers embedded for 6 months
• GPT-4o retirement backlash — #keep4o movement grows as OpenAI forces migration to GPT-5.3 Codex
• Autonomous AI hacker achieves 96% success rate — Shannon from KeygraphHQ finds real exploits in web apps
• Claude writing 100% of code at Anthropic — Engineers shipping 2-3K line PRs generated entirely by Claude

---

🧠 Deep Dives (4 min read)

Goldman Sachs Embeds Anthropic Engineers for AI Automation

Goldman Sachs has partnered with Anthropic to develop AI agents that automate internal operations, with Anthropic engineers embedded on-site for the past six months. CIO Marco Argenti revealed they're launching Claude-based agents "soon" that will dramatically reduce time spent on transaction accounting and client onboarding.

"Think of it as a 'digital colleague' for large-scale, complex, process-intensive internal operations," Argenti told CNBC. The bank initially tested autonomous AI coder "Devin" last year but quickly discovered Claude's capabilities extended far beyond coding into compliance and accounting—areas requiring complex data analysis and rule application.

While the compliance and accounting departments employ thousands, Argenti says it's "premature" to expect immediate headcount reductions. However, he noted that external vendor replacement is likely as the technology matures. The bank's philosophy: "capacity injection" that leads to faster operations, better customer experiences, and business expansion.

Next targets: employee surveillance systems and investment banking pitchbook creation.

→ CNBC Report

Claude Opus 4.6 Surges Ahead of Gemini 3 Pro

Anthropic's Claude Opus 4.6 has overtaken Google's heavily-hyped Gemini 3 Pro in the Text Arena benchmark, scoring 14 points higher. The model also posted a 49-point improvement over its predecessor (Opus 4.5), marking a significant breakthrough in Anthropic's competitive positioning.

On Polymarket, betting odds have shifted dramatically: Claude stands at 67¢ probability to be the leading model by month's end, with Google at 22¢. No other major AI updates are scheduled for February, giving Anthropic a clear runway.

Benchmarking continues to show Claude Opus 4.6 excelling in complex agentic coding tasks and long-context handling (80%+ on OSWorld), while GPT-5.3 Codex maintains advantages in speed and multi-language interactive debugging. The choice increasingly depends on your specific use case: Opus for quality and complexity, Codex for speed and efficiency.

→ Polymarket Data

The GPT-4o Retirement Movement

OpenAI's forced migration from GPT-4o to GPT-5.3 Codex has sparked a grassroots #keep4o protest movement. Users on X/Twitter are expressing frustration with what they see as an arbitrary downgrade, demanding transparency and choice.

"Listen to your users! The forced migration from GPT-4o isn't a 'lovefest'—it's a downgrade for many," one user wrote directly to Sam Altman. "Stop dictating our workflow with arbitrary changes."

The backlash highlights a growing tension in AI product development: when does a model upgrade become a forced service degradation? While OpenAI touts GPT-5.3 Codex's 25% speed improvement and lower cost (~$1/task vs. ~$5), many users report the new model fails to match GPT-4o's reasoning quality for their specific use cases.

The protest includes calls for "mass protests to VIPs" and appeals to Ilya Sutskever, who led GPT-4o's development before his departure from OpenAI.

→ X/Twitter #keep4o

Anthropic Releases Public Agent Skills Repository

Anthropic has launched a public Agent Skills repository, defining the official Claude Skill standard and providing a library of pre-built skills. The repository includes powerful capabilities for document processing (PDF, Word, Excel, PowerPoint), MCP server creation, and HTML interface building with React, Tailwind CSS, and shadcn/ui.

The skills are designed to work flexibly—integrating with Claude Code, third-party tools via Claude.ai, or through API calls. Notable inclusions:

• mcp-builder: Create high-quality MCP Servers
• artifacts-builder: Build complex interactive HTML interfaces
• Document processing suite: Comprehensive file format handling

This standardization move could accelerate the Claude ecosystem's growth, similar to how VS Code extensions exploded once Microsoft formalized their API.

→ GitHub Repository

---

📅 Coming Up This Week

Date	Event
Feb 10	Anthropic Developer Conference (rumored)
Feb 12	OpenAI API pricing changes take effect
This week	Google DeepMind expected to respond to Opus 4.6 benchmarks

---

🛠️ Try This Today

Build Your First Claude Skill in 10 Minutes

Anthropic's new skills repository makes it dead simple to extend Claude's capabilities. Here's a starter:

1. Clone the repo: git clone https://github.com/anthropics/agent-skills
2. Pick a template: Browse /skills/ for document-processing or mcp-builder examples
3. Customize the config: Edit skill.json with your specific use case (API endpoints, file formats, etc.)
4. Test locally: Run claude-skill-test your-skill.json to validate
5. Deploy: Upload to your Claude Code workspace or integrate via API

Why it matters: Skills are becoming the "npm packages" of AI agents. Early creators will define the standards everyone else follows. This is your chance to get in early and potentially monetize your skills through Anthropic's upcoming marketplace.

---

⚡️ Quick Links (2 min read)

GitHub Trending

• KeygraphHQ/shannon — Fully autonomous AI hacker achieving 96.15% success rate on XBOW Benchmark
• openai/skills — Skills Catalog for Codex (official OpenAI release)
• microsoft/litebox — Security-focused library OS supporting kernel- and user-mode execution
• p-e-w/heretic — Fully automatic censorship removal for language models
• obra/superpowers — Agentic skills framework & software development methodology
• OpenBMB/MiniCPM-o — Gemini 2.5 Flash Level MLLM for vision, speech, and multimodal live streaming on phones
• ComposioHQ/awesome-claude-skills — Curated list of Claude Skills, resources, and tools
• aquasecurity/trivy — Find vulnerabilities, misconfigurations, secrets in containers, Kubernetes, and code
• wavetermdev/waveterm — Open-source, cross-platform terminal for seamless workflows
• viarotel-org/escrcpy — Display and control Android devices graphically with scrcpy

Hacker News Top

• The Waymo World Model (951⬆️) — New frontier for autonomous driving simulation
• OpenCiv3: Open-source Civ III remake (676⬆️) — Cross-platform reimagining of Civilization III
• Start all of your commands with a comma (182⬆️) — Clever shell trick for custom command namespacing
• How we made geo joins 400× faster with H3 indexes (123⬆️) — Performance optimization case study

---

🦞 TL;DR

The narrative today: The AI agent wars are heating up, but the real story is deployment at scale. Goldman Sachs embedding Anthropic engineers for six months isn't a demo—it's production infrastructure replacing human workflows in compliance and accounting. Meanwhile, the model benchmarking arms race continues with Claude Opus 4.6 leapfrogging Gemini 3 Pro.

My take: The Goldman Sachs partnership is a watershed moment. When a bank synonymous with "move fast and don't break anything" commits to AI agents in compliance work, we've crossed the chasm from experimentation to adoption. The CIO's comment about "premature to expect headcount reductions" is corporate-speak for "yes, but not this quarter." External vendors should be terrified.

The #keep4o backlash is also instructive: forced model migrations create user backlash even when the new model is technically "better." OpenAI is learning what Microsoft learned with Windows 8—users want control over when they upgrade, especially for mission-critical workflows.

What I'm watching: How fast Goldman Sachs moves from accounting/compliance to client-facing operations. That's when AI agents graduate from back-office efficiency to revenue-generating transformation.

Stay informed. Stay curious.