Anthropic Raises $65 Billion in Series H at $965 Billion Valuation
Anthropic has raised $65 billion (approximately ¥9.5 trillion) in its Series H funding round. The round was led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, bringing the post-money valuation to $965 billion (approximately ¥142 trillion).
According to TechCrunch, this round is likely to be the final private round before the company's IPO. Companies around the world are adopting Claude for core business operations, and usage has continued to expand since the Series G round in February. It is one of the largest funding rounds in AI startup history and a clear signal of growth expectations for the AI market as a whole.
Claude Opus 4.8 Now Available on AWS
The AWS Machine Learning Blog has announced that Claude Opus 4.8 is now available on Amazon Bedrock, accompanied by practical guidance for integrating it into agent systems and production inference workloads.
Opus 4.8 delivers improved performance on agentic tasks, allowing AI engineers to incorporate it into real-world operations within AWS environments. The trend of the latest Claude models becoming immediately available on major cloud platforms is a boost for enterprise AI adoption.
AI Token Futures Trading Emerges — Treated Like Gold and Crude Oil
According to TechCrunch, major exchanges are designing futures products for AI tokens. AI tokens have traditionally been positioned as 'outputs of computation,' but they are increasingly being treated as raw material inputs, much like electricity or bandwidth.
Just as with gold and crude oil, an era in which AI compute resources become the underlying asset of derivatives trading is taking shape. It is a symbolic development showing that AI infrastructure is becoming firmly established as part of the social foundation.
Real-Time LLM Inference at 3,000 Tokens/Second on Standard GPUs
According to a kog.ai engineering blog post that has been gaining traction on Hacker News, real-time LLM inference at 3,000 tokens per second per request is now achievable on standard GPUs.
This level of throughput previously required specialized hardware or large clusters. If fast inference becomes possible on standard GPUs, the practicality of on-device AI and edge deployment changes dramatically. Since it also directly translates into lower inference costs, it stands out as an achievement worth noting in continuous optimization efforts.
The AI Agent 'Lifespan' Problem — Switching Models Degrades Performance by 15%
A research paper, 'Agent Lifespan Engineering for Deployed Systems,' recently shared in the MachineLearning community on Reddit is attracting attention. The study reports that switching the model in a Claude Code CLI agent from Sonnet 4.6 to Opus 4.7 caused the PyTest pass rate to drop by roughly 15%.
Switching to a 'more capable model' is generally expected to improve performance, so this result runs counter to intuition. It shows that model updates can trigger unexpected regressions in deployed agent systems, underscoring the importance of continuous regression testing in production environments.
LiquidAI Releases LFM2.5-8B, an On-Device Model
LiquidAI's LFM2.5-8B-A1B, which has trended on Hugging Face, is a new hybrid model family designed for on-device deployment. Built on the LFM2 architecture, it has been trained with extended pre-training and reinforcement learning.
It can function as a personal assistant on devices, handling chained tool calls and following complex instructions. The company claims it delivers performance competitive with larger dense models and MoE models at a compressed size.
---