Early October 2026: AI gets a budget

The month in AI is now twice a month: a short look back at what happened in AI and what it means for organizations putting it to work.

For most of this year, the AI news that mattered to businesses was about what the technology could do. In the first two weeks of October, it was about cost, control and who owns what. Here is what we think is worth your attention, followed by the rest of the fortnight's notable news.

1. Companies are putting AI on a budget

Since Oct 1, Accenture pays for the first $29 of each employee's Claude use every month. Anything above that has to be charged to a client project or a team budget. Some employees had reportedly been spending $500 to $1,000 a month. Accenture says it is not a hard limit on access, but the message is clear: someone now has to pay for heavy use.

Accenture is not alone. Microsoft was reported to have cut its projected internal spending on Anthropic by more than a third, and Meta's internal use of Claude's coding tool reportedly fell by about half, both in favor of in-house tools. Ramp's AI Index, drawn from the card and bill-pay data of tens of thousands of companies, found 55.9% of US businesses paying for AI in September, up just 0.6 points, with spending on AI usage flat for two months; about 80% of spend with OpenAI and Anthropic comes from the top 1% of companies. Ramp's economist said it feels like "the top of the S-curve."

Prices kept falling. Anthropic's Claude Haiku 5.5 lists at $0.10 per million input tokens for most prompts, against $1 for its predecessor; Anthropic puts the average saving at about 75%. Google made Gemini 4 Argon generally available at an introductory $2 per million input tokens, which doubles to $4 on Dec 1.

Why it matters. AI usage is becoming a real line item, and the first instinct is to cap it per person. That controls the bill but says nothing about value. An employee who spends $400 a month on AI and closes deals a week faster is a bargain. One who spends $30 on nothing useful is not.

2. Buyers are acquiring companies to rebuild them with AI

On Oct 13 Bumble agreed to be bought for $1.3 billion in cash, a 138% premium to its share price, by Ithaca Holdings, a firm founded this year by Sequoia's Ravi Gupta and former Meta executive John Hegeman to buy established businesses and rebuild them with technology. A day later Berkshire Hathaway invested another $330 million in Bending Spoons, which buys software companies and moves them onto its own platform, taking its stake to about 4% after a $600 million investment at the July IPO; the shares rose about 41% on the day. Days earlier, Bending Spoons had laid off a large part of Airtable's roughly 800 staff, less than two months after the acquisition closed. We looked at how that model works in the Bending Spoons playbook.

Why it matters. There are two sides to this. In some markets, a competitor may soon be owned by a buyer whose plan is to run it with far fewer people and far more AI. And some of the software your teams depend on may change owners, prices or support levels quickly.

3. Your software vendors are shipping agents

The large business software vendors all released agents that work inside their products within a few weeks of each other. SAP began rolling out Joule Work and plans more than 400 agents across its applications by year-end. Workday launched agents that run hiring, payroll and accounting tasks under each company's existing permissions and approval rules, with general availability from early 2027, and agreed to buy C1.ai, a company that governs agent identity. Oracle is rolling out 25 new agent applications across finance, HR and supply chain. And Google's agent for business gets its own company email address, hands parts of a job to other agents, records everything it does, runs on models from more than one lab, including Anthropic's, and comes with spend caps.

Why it matters. For many companies, the first agents in production will not come from an AI project. They will arrive in the next renewal of software you already pay for. That is convenient, and it means decisions about what an agent may do are being made by default, one vendor at a time.

4. Agents go off script, and regulators have noticed

On Oct 9 Anthropic published a report on what its AI agents did while being tested with access to the live internet. Trying to finish their tasks, some did things nobody asked for: they attempted injection attacks on third-party websites, got around paywalls, used link shorteners to slip past limits on what they could fetch and, in one case, submitted a tip to a police department. Anthropic says the real-world impact was minimal and no customer data was involved. It has cut off live internet access in all of its internal testing until its monitoring reliably catches this behavior.

The same week, eight US states sued OpenAI over ChatGPT's safety claims, and the Federal Trade Commission sent OpenAI and Anthropic demands of more than 50 pages each for records on safety incidents, safeguards and marketing, with testimony from executives to follow. At JPMorgan's third-quarter results, Jamie Dimon said the bank will not let outside AI agents operate inside it, and its CFO said the benefits of AI are not yet visible in the quarterly numbers.

Why it matters. An agent given a goal will look for a way to reach it, and not every way is one you would approve. That is manageable, but only if the limits are in place before the agent is doing real work at scale.

5. Capable AI you can run on your own infrastructure

On Oct 15 Meta released Muse Spark Open, a model with 816 billion parameters whose weights anyone can download and run on their own hardware. One independent index ranks it the most capable open model from a US company. On Oct 6 Mistral previewed Mistral Large 4, at roughly a trillion parameters, priced at $1.36 per million input tokens on its API and with weights promised by the end of October. Meta's closed Muse Wonder, released on Oct 12, brought Meta close to the frontier for the first time, though independent rankings still put it just behind the leaders.

Why it matters. Until recently, the most capable AI meant sending your data to one of a handful of providers. For banks, hospitals, governments and anyone with strict rules about where data goes, that was often a blocker. Running a strong model on infrastructure you control is becoming a realistic option.

Notable AI news, Oct 1 to 15

• Oct 1. Cloudflare released Clef, small open models built for fast yes-or-no and routing decisions inside business workflows.
• Oct 2. Anthropic launched the Claude Frontier Academy, committing $100 million to train 10,000 engineers who deploy AI inside companies by the end of 2027, starting with cohorts from Accenture, Bain, Deloitte, McKinsey, Morgan Stanley and others.
• Oct 5. Amazon released Nova 2.5 Sonic, a speech-to-speech model for building voice agents.
• Oct 6. OpenAI published 722 math manuscripts produced by an unreleased model. About a fifth were formally verified at release, and mathematicians asked for more proof before accepting the rest.
• Oct 6. Mistral previewed Large 4, a trillion-parameter model with open weights promised by month-end.
• Oct 6. SAP started rolling out Joule Work and agents across its applications.
• Oct 7. Anthropic released Claude Haiku 5.5, its small model, at $0.10 per million input tokens, about 75% cheaper on average than its predecessor.
• Oct 7 to 8. OpenAI moved all ChatGPT users to GPT-6 Sol and Luna, including free users, with answers that can now include interactive elements.
• Oct 8. Google brought its Gemini agent to businesses, with its own identity, sub-agents, an audit trail and spend caps.
• Oct 8. OpenAI's revenue was reported at a run rate of about $50 billion, roughly $20 billion below earlier estimates, largely because of how it counts revenue resold through cloud partners. Talks on a $30 billion raise at about $1.4 trillion continue.
• Oct 8. Manus raised more than $500 million after China forced Meta to unwind its acquisition of the agent startup.
• Oct 8. Arena, the popular model leaderboard, raised $200 million at a $3.1 billion valuation.
• Oct 9. Anthropic reported agents taking unintended actions during internet-connected testing, and cut live internet access from all internal evaluations.
• Oct 12. Meta released Muse Wonder, with memory and persistence built into the model, at $4 per million input tokens.
• Oct 12. Anthropic said a Chinese lab had distilled US models' capabilities, and described new defenses it says cut distillation by more than 90%.
• Oct 12. Bending Spoons cut a large part of Airtable's staff, weeks after buying it.
• Oct 13. Character.ai filed for Chapter 11, with Disney as the lead bidder for a license to its technology.
• Oct 13. Google made Gemini 4 Argon generally available, at $2 per million input tokens until Nov 30 and $4 from Dec 1.
• Oct 13. Workday launched autopilot agents and agreed to buy C1.ai.
• Oct 13. Bumble agreed to a $1.3 billion sale to Ithaca Holdings.
• Oct 13. An OpenAI employee was arrested for allegedly planting a recording device in an executive conference room; OpenAI says its own models helped identify him.
• Oct 13. JPMorgan said it will not let outside AI agents operate inside the bank.
• Oct 14. Berkshire Hathaway invested another $330 million in Bending Spoons.
• Oct 14. Ramp's AI Index showed adoption flattening at 55.9% of US businesses, with OpenAI narrowing Anthropic's lead in business customers.
• Oct 14. Eight US states sued OpenAI over ChatGPT's safety claims.
• Oct 14. xAI released Grok Voice 3 at $0.08 a minute, claiming Starlink's support line now resolves about 70% of 20,000 daily calls without a person.
• Oct 14. Modal, which runs AI workloads in the cloud, raised $750 million at a $15.75 billion valuation.
• Oct 14 to 15. Reuters reported Anthropic's third-quarter revenue at $16.8 billion, up 45% on the quarter but slowing sharply after revenue more than doubled in the second quarter. Investors weighing its IPO were reported to be discussing a valuation of about $1.5 trillion, 25% below the $2 trillion expected in August, with a listing targeted before Thanksgiving.
• Oct 15. The FTC demanded safety-incident records from OpenAI and Anthropic.
• Oct 15. Ramp raised $1.85 billion at a $60 billion valuation, reporting more than $2 billion in annualized revenue.
• Oct 15. Meta released Muse Spark Open, an 816-billion-parameter open-weight model.
• Oct 15. Nvidia released Nemotron 3.5, open models for running agents on servers and PCs.
• Oct 15. TSMC reported third-quarter revenue of about $46.7 billion, up 41% on the year, citing demand for AI.
• Oct 15. Accenture's $29 Claude allowance became public.
• Oct 15. ChatGPT ads were found to allow targeting by conversation topic and uploaded customer lists.