Google Built One Agent To Run Your Whole Stack

The Big Story

Anthropic shipped Claude Haiku 5.5 on October 7 across its own platform, AWS, Google Cloud, and Azure, and the news is the price. For prompts up to 100,000 tokens it lists at $0.10 per million input tokens and $0.50 per million output, down from $1 and $5 on Haiku 4.5, which Anthropic frames as about 75% cheaper on average and closer to 90% on short prompts. It also cut the Sonnet 5.5 cache-read price in half, to $0.10. The catch sits at the 100K mark, where rates jump five times, to $0.50 and $2.50. For anyone building agent swarms or high-volume classification, this is the number that makes routing cheap work to a cheap model worth the engineering. For anyone running long-context agents, that price cliff is a line to design around.

OpenAI and Atlassian expanded their partnership this week to bring OpenAI frontier models into Rovo and the agents that run across Atlassian's platform, according to week-of-October-9 roundups. For the many enterprises whose engineering and project work lives in Jira and Confluence, this is the same pattern Salesforce showed at Dreamforce last week: the system of record stays Atlassian, while the model doing the reasoning inside it becomes OpenAI's. The buyer gets capability without a migration. The buyer also gets another vendor's model reading its project data by default, which makes this a data-governance review before it is a feature toggle. Worth confirming scope against Atlassian's own release notes before anyone turns it on broadly.

A National Bureau of Economic Research working paper built on the Duke CFO survey found that 44% of finance chiefs at 750 US firms plan AI-related job cuts this year, which the authors scale to roughly 502,000 roles across the economy in 2026, about nine times the 55,000 AI-attributed cuts recorded in 2025. Set that intent against the record so far: Challenger, Gray & Christmas has linked nearly 88,000 actual 2026 losses to AI. The gap between planned and realized is the story. Executives are budgeting headcount reductions against a capability their own deployments have mostly not delivered yet, which makes the plan a bet on next year's agents rather than this year's results. If you are the one asked to make the math real, the pressure is now a line item.

Microsoft merged its consumer and workplace Copilot into a single, business-first application, per week-of-October-9 reporting. The move ends the split between the Copilot employees used at home and the one IT certified at work, and it does so by making the work version the default shape of the product. For admins, consolidation cuts one axis of shadow-AI confusion, since there is now one app to govern rather than two that drift apart. It also concentrates more of the daily surface under a single Microsoft identity and policy layer, which is convenient until you want a tool that is not Microsoft's sitting next to it. Check what the merge changes about default data handling before it rolls to your tenant.

Nous Research closed a $90 million Series B at a $1.5 billion valuation to push its Hermes agents into companies through a new Hermes for Businesses offering, with reported participation from Nvidia, Menlo Ventures, and others. The Wall Street Journal pegged its annualized revenue near $36 million by mid-September. The pitch is open-weight agents a company can run on its own data, with the privacy and control that implies, against the hosted-model default. It is a small revenue base behind a large valuation, so the round is priced on the thesis that enterprises will pay for agents they can inspect and self-host. For a CIO weighing lock-in, the open-weight lane is now funded well enough to be a real option on the shortlist.