I. Funding & Industry Landscape: The Trillion-Dollar Club Expands

Anthropic's $965B Valuation: AI's "Apple Moment"

Anthropic completed a $65B Series H round led by Altimeter Capital, reaching a post-money valuation of $965B with annualized revenue exceeding $47B. This valuation approaches the world's top 10 tech companies by market cap. Anthropic's revenue growth is primarily driven by enterprise deployments — Claude's paid usage in coding agent scenarios has surged, with enterprise plans shifting from fixed discounts to usage-based billing.

DeepSeek's STAR Market IPO: China's AI Capitalization Accelerates

DeepSeek plans to file for a STAR Market IPO immediately after completing approximately $50B in funding — the first Chinese AI foundation model company to clearly define an IPO path. Combined with Alibaba Cloud being named the leader in agent AI by Omdia and Qwen3.7-Max topping OpenRouter's popularity chart, Chinese AI's global presence continues to rise.

OpenRouter's $113M Series B: Model Aggregation Layer Validated

OpenRouter secured a Series B led by CapitalG, confirming that "model routing" as a middleware layer has become essential. As model count explodes, developers need a unified gateway to switch, compare, and reduce costs.

II. Model Releases: Agent Capabilities as Core Selling Point

Claude Opus 4.8: Incremental Upgrades, Coding is King

Anthropic released Claude Opus 4.8, positioned as an upgrade to 4.7 with improvements in coding, agent skills, and reasoning at the same price. The simultaneously launched "Dynamic Workflows" feature enables Claude Code to run dozens to hundreds of sub-agents in parallel within a single session, handling cross-codebase bug hunting and large-scale migrations.

Grok Build 0.1: xAI Enters the Coding Agent Arena

xAI's Grok Build is specifically trained for agent coding tasks, supporting web development, debugging, and MCP with inference speeds exceeding 100 tok/s. This marks xAI's move beyond chat models into direct competition with Cursor/Claude Code.

Step 3.7 Flash: Chinese Models Take the Efficiency Route

Step 3.7 Flash (198B MoE) ranked first in ClawEval and SimpleVQA Search benchmarks, focusing on agent workflow efficiency. Chinese models are shifting from "catching up on general capabilities" to "optimizing for specific scenarios."

III. Products & Tools: Agent Ecosystem Taking Shape

This week's product launches reveal a clear trend: AI is moving from "chat box" to "workbench."

MCP (Model Context Protocol) is becoming the de facto standard for connecting AI with external tools.

IV. Research Frontiers: New Paradigms for Agent Training & Evaluation

NVIDIA Polar Framework: Breakthrough in Agent RL Training

NVIDIA open-sourced Polar, an agent reinforcement learning framework that doesn't require rewriting existing agent frameworks (Codex CLI, Claude Code, etc.), connecting to GRPO training by placing agents at model API boundaries. Experiments show Qwen-based models improved Codex benchmark scores by 594.74%.

SIA Framework: AI Recursive Self-Improvement

hexoai's open-source SIA (Self-Improving AI) framework demonstrates that AI agents can not only optimize external workflows but also directly update their own model weights through task feedback — another proof of "AI training AI."

KPop: Stable RL Training for Large-Scale MoE Models

The KPop method pushed Ring-2.6-1T past 76 on SWE-bench Verified, using adaptive masking to replace fixed-ratio masks, solving stability issues in large-scale MoE model RL training.

V. Trend Signals

Key Numbers This Week

MetricValueSignificance
Anthropic Valuation$965BApproaching trillion, first AI company at this level
Anthropic Annual Revenue$47BEnterprise AI payments at scale
DeepSeek Funding Valuation$50BAmong China's highest AI unicorn valuations
Cognition Annual Revenue$4.92BAgent lab commercialization validated
Polar Framework Improvement594.74%Massive potential of agent RL training
Developer Weekly Code Output8.6K linesProductivity doubled with AI assistance

Overall Analysis

The most significant signal in AI this week: capital scale and technical competition have simultaneously entered a new magnitude.

From the capital side, Anthropic's $965B valuation, DeepSeek's $50B IPO push, OpenRouter's $113M Series B — the market is no longer debating "whether AI has value" but competing for "who becomes the next trillion-dollar platform." Anthropic and OpenAI simultaneously shifting from "subsidized acquisition" to "usage-based billing" indicates coding agent usage has grown large enough to support massive revenue — PMF confirmed.

From the technology side, focus has shifted from "bigger, stronger models" to "how agents land in production." Claude Opus 4.8's Dynamic Workflows, NVIDIA Polar's 594% improvement, SIA's recursive self-improvement — these advances point in the same direction: AI coding capabilities can continuously evolve through automated RL training, no longer dependent on human annotation. Anthropic's strategy is particularly clear: no pursuit of shocking launches, but high-frequency iteration, with each upgrade focused on "making daily developer use smoother."

From the product side, "embedded AI" has become mainstream — not making users find AI, but making AI appear in existing workflows. For developers, API costs will become an increasingly important selection factor, and model aggregation platforms (like OpenRouter) are being validated by capital for this reason.

In one sentence: In Q2 2026, the AI industry has officially moved from "technology validation" to "commercial scaling," with coding agents as the first confirmed killer application.

Next Week Outlook