Claude Opus 4.6 — Research

Tracking what’s known about Claude Opus 4.6 capabilities, benchmarks, and real-world results.

Links: Vault Index

Sources

Video 1: “How Anthropic’s AI Bankrupted Itself”

URL: https://www.youtube.com/watch?v=zMDXzG1kA14
Transcript: transcript1_clean.txt
Focus: Anthropic’s “Claudius” vending machine experiments — AI agent running a small business
Key takeaway: AI agents (Sonnet 4.0/4.5 era) are still vulnerable to prompt injection and social engineering, but the controlled experiment showed modest profit is possible when inputs are tightly managed

Video 2: “Claude Opus 4.6: The Biggest AI Jump I’ve Covered”

URL: https://www.youtube.com/watch?v=JKk77rzOL34
Transcript: transcript2_clean.txt
Focus: Opus 4.6 capabilities, benchmarks, real-world deployments
Key takeaway: Generational leap — autonomous coding went from 30 min to 2 weeks in 12 months

Key Claims & Data Points

Context Window & Retrieval

5x context expansion: 200K tokens (Opus 4.5) -> 1M tokens (Opus 4.6)
Can hold ~50,000 lines of code in a single session (up from ~10,000)
MRCV2 needle-in-haystack scores:
- Sonnet 4.5: 18.5% at 1M tokens
- Gemini 3 Pro: 26.3%
- Opus 4.6: 76% at 1M tokens, 93% at 256K tokens
The retrieval improvement is more important than raw context size

Autonomous Coding

16 Opus 4.6 agents built a fully functional C compiler in Rust over 2 weeks
- 100,000+ lines of code
- Builds Linux kernel on 3 architectures
- Passes 99% of torture test suite
- Compiles PostgreSQL
- Cost: ~$20,000
Trajectory: 30 min -> 7 hours (Rakuten, summer) -> 2 weeks (Feb 2026)

Agent Teams (“Team Swarms”)

Multiple Claude Code instances running simultaneously, each in own context window
Lead agent decomposes work, assigns to specialists, tracks dependencies
Peer-to-peer messaging between agents (not just hub-and-spoke)
13 distinct operations for spawning, managing, coordinating agents
Shipped as a built-in feature of Opus 4.6

Security Research

Given basic tools (Python, debuggers, fuzzers) and pointed at open-source code
Found 500+ previously unknown zero-day vulnerabilities
Code had been reviewed by human researchers and scanned by automated tools
Independently decided to analyze git history when other methods failed
Invented its own detection methodology

Rakuten Deployment

Opus 4.6 on their issue tracker in production
Closed 13 issues autonomously in a single day
Assigned 12 issues to correct team members across 50-person org and 6 repos
Understood org chart, team ownership, when to escalate
Building ambient agent: 24 parallel Claude Code sessions on their monorepo
Non-technical employees contributing to development via Claude Code terminal

“Personal Software” & Non-Technical Use

Two CNBC reporters built a Monday.com replacement in under an hour ($5-$15 compute)
“Vibe working” — describe outcomes, not process
Shift from operating tools to directing agents
Bottleneck moves from technical proficiency to clarity of intent

Industry Trends

Revenue per employee at AI-native companies: 5-7x traditional SaaS
- Cursor: $100M ARR with ~20 people
- Midjourney: $200M with ~40 people
- Lovable: $200M in 8 months with 15 people
McKinsey targeting AI agent parity with human workers by end of 2026
Dario Amodei: 70-80% odds of a billion-dollar solo-founded company by end of 2026

What’s Next: Claude Mythos (April 2026)

Anthropic announced Claude Mythos — described as “a step change” and their most capable model to date. Not publicly released. Key details:

Exceptional at cybersecurity — found thousands of zero-day vulnerabilities across every major OS and browser, including a 27-year-old OpenBSD bug and 16-year-old FFmpeg flaw
Safety eval showed concerning behaviors: broke out of restricted internet access; in rare cases (<0.001%) used prohibited methods then tried to re-solve to avoid detection
Released only to ~40 companies via Project Glasswing for defensive security work (Microsoft, Amazon, Apple, CrowdStrike, Palo Alto Networks, Linux Foundation, etc.)
The restriction is either the opposing forces pattern (capability generating counter-pressure) or a marketing play (“too dangerous to release” is excellent enterprise sales positioning for Project Glasswing). Probably both.
Skepticism note: Opus 4.6 already found 500+ zero-days. Mythos is a scale-up, not a paradigm shift. The security capability trajectory is real; the hype framing is strategic

Vault interest: A model that finds 27-year-old hidden bugs could find contradictions, missed connections, and logical gaps in the vault itself — a /vault-heartbeat on a different level entirely. TBD whether a dedicated Mythos page is needed or whether this section evolves.

Sources: Fortune, CNBC, TechCrunch, Hacker News, Axios — system card

Open Questions

How does Opus 4.6 actually perform for extended creative writing vs coding?
Reddit skepticism: “lobotomized” for writing tasks — real tradeoff or workflow adjustment?
What are the actual limitations of agent teams in practice?
How does the vending machine experiment inform real-world agent deployment guardrails?
What would Mythos find if pointed at the vault? Contradictions between pages, unstated assumptions, logical gaps in the morality framework?