Anthropic just made capable agents cheap enough to leave running all day
Claude Sonnet 5 shipped Tuesday as the new default for free and Pro users, and the pitch is blunt: agent-grade planning, browser and terminal use, and long autonomous runs at a midsize price. Intro pricing is $2 per million input tokens and $10 output through August 31, then $3 and $15, still under Opus 4.8, GPT-5.5 and Gemini 3.1 Pro. Work that needed a frontier model a few months ago now runs on the cheap tier. If you priced an agent workflow as too expensive earlier this year, the math changed, so re-run it before you rule it out.
Source: Original source
OpenAI raised $122 billion and the number that matters is the valuation
OpenAI closed a fresh round at $122 billion committed, pushing its post-money valuation to $852 billion, days after previewing the GPT-5.6 Sol line built to split work across subagents for longer autonomous jobs. The capital is aimed squarely at compute. For everyone downstream, the read is that the frontier labs are raising at a scale that assumes agents, not chat, are the product. Budget and plan accordingly, because the tools you build on are being funded to run work, not answer questions.
Source: Original source
The agent question flipped from "is this real" to "what gets automated first"
Gartner now projects AI agent software spending hits about $206.5 billion in 2026, up 139% from last year, the fastest-growing slice of enterprise software. That figure is less a prediction than a sign that budgets already moved. The useful question for an operator is no longer whether agents work, it is which process in your own shop is the first honest candidate: the repetitive, rules-heavy, high-volume one where a mistake is cheap to catch. Pick that, prove it, then widen.
Source: Original source
Sovereign AI stopped being a slogan and started signing power contracts
NAVER and NVIDIA committed to build Korea's national AI infrastructure on NVIDIA's DSX platform, opening at 55 megawatts with a stated path to gigawatt scale, feeding NAVER's HyperCLOVA X models and an agent platform launching in Korea later this year. Countries are now buying their own model stacks measured in power, not licenses. If you sell into regulated or government-adjacent markets, expect "where does this run and who controls it" to become a real line in the deal, not an afterthought.
Source: Original source