The Superintelligent Way · Monthly round-up
Prices fell hard, one big lab refused to ship its own model, and the skills that pay changed. Here is the month in five parts: the new models, who is really using them, the skills that pay, what changed for builders and what changed for marketers. Each part has the key numbers, what they mean for you, and the sources.
At a glance
Part 1 · Models
Google's new top model reached the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers on 13 October, and developers can use it through the Gemini API. The API costs $2 per million input tokens and $10 per million output tokens until 30 November, then $4 and $20.
Sonnet 5.5 arrived on 28 September and Haiku 5.5 on 7 October. The big story is the small model. Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, for prompts up to 100,000 tokens. Haiku 4.5 cost $1 and $5.
Ten times cheaper, and better. On OSWorld, a test where the model uses a computer the way a person does, Haiku 4.5 scored 15.7%. Haiku 5.5 scores 72.4%. Anthropic's own numbers
Sonnet 5.5 kept Sonnet 5's price of $2 and $10, and on Terminal-Bench 4.0, a test of coding in a terminal, it scored 70.6% against Sonnet 5's 10.3%. Anthropic's own numbers
Output price, per million tokens
Claude's small model, last generation against this one
Anthropic list prices. Input fell the same way: $1 to $0.10 per million tokens, for prompts up to 100,000 tokens.
Using a computer like a person
OSWorld 2.1 score, % of tasks completed
Anthropic's own numbers, from its published tables. The Haiku and GPT-6 Luna scores use an offline subset of the test.
On 28 September OpenAI said it would not release its next model, GPT-6.1 Astra, because it "didn't quite meet the bar". The concern was safety, not intelligence. Saachi Jain, its head of safety systems, said it fell short on "scope and authorization, and how it communicates back to the user about the type of work it's done". No new release date was given.
Three open models you can run yourself came out this month.
Part 2 · Adoption
Ramp tracks what companies pay for on their cards. In September, 55.9% of US businesses paid for AI tools, and exactly half of small businesses. Growth is slowing: the share rose by 0.62 points in a month, and Ramp's economist suggests adoption may be near the top.
Share of US businesses paying for AI, by sector
September 2026, % of businesses
Ramp AI Index (card and bill-pay data), published by the St. Louis Fed's FRED.
That's the gap. Gartner surveyed 2,002 CIOs and tech executives. Only 13% said AI had delivered significant or transformational value, and about 1 in 5 AI use cases succeed.
Spending continues because CEOs expect it to pay back. In KPMG's survey of 400 US CEOs, 84% call AI a top investment priority for the next three years, and 68% expect a return in one to three years.
A bank, with strict rules, that started with email. In its Global Markets division, Claude now handles about 120,000 client emails a day. Its internal knowledge assistant is used by more than 16,000 staff, and the bank aims to have half its developers on Claude Code by the end of 2026.
Part 3 · Skills
Workday looked at hiring at more than 550 employers. Demand for basic prompting is down 25% from its January peak. Demand for people who build with AI (tools, automations, agents that do a real job inside a business) is up 51%.
Change in demand
% change; the measures differ, see the note
Skills: Workday job requisitions at 550+ employers (prompting since its January 2026 peak; hands-on AI Sep 2025 to Jul 2026). Job posts: CyberSN, US postings January to September 2026 against 2025, via SecurityWeek.
That's someone who makes AI work inside a real company. In 2023 there were fewer than 200 US job posts for it. From January to September 2026, more than 7,300, at a median advertised salary of $195,456.
Anthropic is putting $100 million into training 10,000 of them by the end of 2027. Companies nominate their engineers; individuals can't apply.
US job posts for forward deployed engineers
Number of posts
Lightcast data, via CNBC, 15 October 2026. Median advertised salary $195,456, up 40% from 2023. The 2023 bar is drawn at its upper bound.
On a CV, "I can prompt" doesn't stand out anymore. "I automated this" does.
Part 4 · Development
Anthropic's Python and TypeScript SDKs can now control a computer and a browser (beta). Form filling, data entry, testing a website. The cheap Haiku 5.5 can run them.
7 October
Copilot now runs agents in a local sandbox that blocks them from your files, your network and your credentials, at no extra charge, in Copilot CLI, the app and VS Code.
9 October
Most reviews are about three times faster: 6 to 10 minutes instead of about 20. You pick Quick, Standard or Deep for each repository.
15 October
Cursor's iPhone app controls agents running on your own computer.
6 October
Microsoft dropped a planned UK data centre at Loughton, run by Nscale, with 23,040 Nvidia GB300 chips, because the grid can't supply the power until the early-to-mid 2030s.
Part 5 · Marketing
It can plan a campaign, make the creative and change targeting and budgets from a chat, and you approve each change. Meta says advertisers who follow its recommendations "typically are seeing 5% more conversions at the same cost per conversion". Meta's own claim
6 October
For the first time, you can mark Search, Shopping, YouTube, Display and other channels High or Low priority, and cap how much of your daily budget Low channels get. It rolls out over the coming months. From February 2027, the current setup ends for Search campaigns with the Display Network on.
14 October
Labelled ads next to generated images start testing in the US this month, for Free and Go users. Ten click-measurement partners, including AppsFlyer and Triple Whale, let you track clicks like any other channel.
5 October
One chat tool to buy ads, and AI creative tools that turn the product photos you already have into video ads. The Ads Agent chat is in an expanded closed beta.
30 September
What the platforms say their AI tools do
% change, each platform's own claim
As claimed by each platform, with no independent check. Amazon's figures are from beta testers with no baseline given. A lower acquisition cost is the improvement Amazon claims.
Global ad spend in 2026, by channel
US$ billions (growth on 2025)
WARC forecast, 14 October 2026, via CFOtech. All channels: $1.34 trillion, +11.9% on 2025.
WARC expects global ad spend to grow 11.9% this year, to $1.34 trillion. Outside China, Alphabet, Amazon and Meta take 59.7% of it: the same three that push these AI tools.
This week
Pick one task you repeat every day, and give it to a cheap model. Then measure what it saves you.
Every item comes from news published between 25 September and 21 October 2026, confirmed at the company's own announcement or a named news outlet. Figures that a company reports about its own product are labelled as its own numbers. Reports we could not confirm are left out.