Four Chinese labs released open-weights coding models in 12 days — all hitting 56–59 on SWE-Bench Pro at under one-third the inference cost of Claude Opus 4.7. The "China is 6–9 months behind" narrative is no longer defensible for agentic coding.
MiniMax M2.7
59
GLM-5.1
58
Kimi K2.6
59
DeepSeek V4
59
Claude Opus 4.7
62
15.9%
GLM-5.1 stock gain on launch day
3x
Cost advantage vs Opus 4.7
12hr
Kimi K2.6 Zig port continuous trace
#coding-agents#china-models#SWE-Bench#agentic
💰 Markets
Anthropic's $900B Moment
Anthropic in talks for $50B raise at $850–900B post-money — up from $380B in February. The step-up: 6x in 3 months. IPO target: October 2026 via Goldman Sachs, JPMorgan, Morgan Stanley.
Feb 2026
Previous raise
$380B valuation · $30B raised
Apr 2026
In talks: new round
$900B valuation · $50B raise
Oct 2026
IPO target
$900B+ public debut
$40B
Google committed
$5B
Microsoft invested
#anthropic#funding#IPO
🏗️ Infrastructure
Agent Infrastructure — The New VC Bet
VCs rotating from "which model wins" to "who deploys models at enterprise scale." Amsterdam's Wonderful raised $150M Series B (Insight Partners) — 900 employees, 30 markets in 8 months.
80%of US VC
● AI deal flow share Q1 2026
● Agent infra (orchestration, compliance)
● Vertical agents (legal, health, compliance)
$150M
Wonderful Series B
900
Wonderful employees (8 mo)
#agent-infra#venture-capital#enterprise
📊 Stanford HAI 2026
12 Key Takeaways
Frontier model training costs hit $490M (Grok 4, July 2025). AI hiring tools showing systemic racial bias — regulatory risk emerging. AI moving from "read path" to "write path" — executing real business operations autonomously.
Grok 4 train
$490M
New scaling
10–100x ↓
Regulatory risk: AI bias in hiring tools
#stanford-hai#scaling-laws#regulation
🚀 Funding Watch
AI Startup Mega-Rounds
Cognition (Devin)
AI coding agent
$25B
follow-on talks
Cursor (Anysphere)
AI code editor
$50B+
raising $2B+
SpaceX / xAI
Merged entity
$1.75T
target IPO
#funding#cognition#cursor#spacex
🧠 Agent Memory
Amnesia Problem → Production Solution
Agent memory is shipping as a first-class product. agentmemory MCP server: zero-config integration with Claude Code, Cursor, OpenClaw, and Hermes. 200K conversations/month at 76% automation for Mobi2Buy.