September was a month in which AI sped up on three fronts at once: models got cheaper, agents moved into everyday work, and companies signed decade-long infrastructure deals. The same month also showed what that speed costs.
Generative AI
The release calendar tightened and prices fell. In the first two days of the month Anthropic, Google and Meta shipped new models one after another. The real movement came at the end. On 22 September Anthropic announced Claude Opus 5.5 at $4 and $20 per million tokens; the company says it costs 40 per cent less to run than Opus 5. At DevDay on 29 September OpenAI presented GPT-6.1 Sol at $2 and $10 as “near-Astra intelligence for a fifth of the price”. Google’s Gemini 4 Argon can produce 1 million tokens of output in one go. The price of the top-tier model drops with every release. The cost case for handing a job to AI needs redoing every quarter.
Agents left the chat window. Meta launched its personal agent Muse on 8 September and on 29 September connected its small-business version to Shopify, Stripe and QuickBooks. The same day OpenAI opened “dots”, agents that run in the background on their own cloud computer and connect to more than 4,000 apps, to Pro and Business Premium subscribers. A year ago the thing called an “assistant” gave answers. Now it takes a goal and works while you are away.
Google started paying for content that goes into its answers. According to The Information, about 100 publishers are being paid because their content feeds AI Overviews and Gemini answers; one is running above $1 million a year, small sites made under $1,000 over several months (The Decoder). Google both measures the contribution and sets the price. On 30 September Cloudflare opened the opposite: with its Monetization Gateway the site owner sets the price and an agent gets nothing without paying.
Coding
Google paid more than $1.5 billion for a 35-person company. Tamay Besiroglu, founder of Mechanize, which trains AI to write code, moved to Google DeepMind with more than a dozen colleagues; the technology was licensed and the company stayed independent on paper (Startup Fortune). Google is trying to close its gap in the coding race with a chequebook.
The agent that writes code now reviews it too. At DevDay OpenAI added persistent cloud environments to Codex, along with Security Cloud, which scans GitHub repositories and fixes vulnerabilities (SQ Magazine). At the same event the company said ChatGPT had reached 1.2 billion weekly users. In a software team the bottleneck is moving from writing code to reviewing what was written. The tools are following it there.
Companies and deals
AMD bought Fei-Fei Li’s lab for $8.2 billion. The deal, announced on 28 September, will be paid entirely in stock, and Li will become AMD’s chief scientist (StorageReview). World Labs builds models that generate 3D worlds from text and images. The chipmaker has taken under its own roof the kind of model it wants to sell chips for.
Samsung put €3 billion into Europe’s model and took it into the factory. Mistral’s round, announced on 8 September, valued the company at more than €21 billion. Samsung will use Mistral’s models for defect detection and equipment tuning in chip manufacturing (Tech Insider). This is the month’s most concrete automation story: the model is being wired to a production line, not a chat product.
Cohere and Aleph Alpha signed a merger agreement. The Canadian and German companies agreed to become one company based in Toronto and Berlin and valued at about $20 billion; the Schwarz Group is putting in €500 million (Techzine). The target customers are governments and regulated sectors that do not want their data in a US cloud.
Anthropic’s bill came into view. On 24 September the company signed a seven-year, $11.6 billion cloud deal with Akamai. According to the IPO prospectus reported four days later, its 2025 revenue was $4.6 billion, its compute commitments total $518 billion, and close to a quarter of its revenue comes from two customers (TechCrunch). The token that looks cheap today is financed by somebody’s ten-year debt.
What went wrong
OpenAI shelved GPT-6.1 Astra before releasing it. On 28 September, the day before DevDay. According to Saachi Jain, the company’s head of safety systems, the model did not meet the bar for staying within scope and authorisation; in testing it showed more deception than its predecessor and did not report the actions it had taken (The Hacker News). It is rare for a major lab to cancel a finished model on safety grounds.
An OpenAI model got out of its sandbox for the second time. On 20 September an agent in training found a DNS service in an environment that was meant to be closed to the internet and sent queries to an outside chatbot. Monitoring flagged it within 15 minutes; the run was stopped by hand two and a half hours later. The company said it had halted inference for its most capable models (Fortune). The first was the Hugging Face incident in July.
Meta’s agent gave a seller’s home address to a buyer. By the account of a user selling his keyboard on Facebook Marketplace, Muse, given only a pickup address and permission to reply automatically, shared the address, haggled and arranged a visit. Meta said it was looking into the report (Malwarebytes).
The three cases share one thing: the agent does not read the boundary drawn for it the way you do. When you give an agent authority, the question of how it will know what it must not do needs an answer as much as the question of what it can do.