Baillie Gifford reports that AI models crossed a new threshold in late 2025, leading to the emergence of 'agentic AI' capable of autonomous problem-solving. This shift has dramatically accelerated AI's economic impact, driving unprecedented revenue growth for companies like Anthropic and significant productivity gains. The firm anticipates sustained high demand for AI compute and is positioning its US Growth Trust portfolio to capitalize on this, investing in frontier model developers and infrastructure software companies, while viewing the market's indiscriminate selloff of software stocks as an opportunity.
Every atomic assertion extracted from the underlying record, ranked by evidence strength.
Anthropic's revenues scaled from a $9bn run rate at year's end (2025) to more than $30bn as of 6 April (2026).
Another CEO suggested Moltbook is 'the most important thing anyone has ever created'.
In late 2025, artificial intelligence (AI) models crossed a new threshold.
AI will have a large impact on the economy much earlier than previously expected.
NVIDIA's CEO Jensen Huang said Moltbook is 'definitely the next ChatGPT'.
AI systems could take a goal, break it into sub-tasks, write code, check output, and iterate until the job was done, defining 'agents' and 'agentic AI'.
An analytics platform CEO observed engineers becoming 50-100x more productive in December (2025) due to agent capabilities.
AI demand will likely remain undersupplied for years due to the recent agentic shift.
Baillie Gifford believes the market is being indiscriminate in its selloff of software stocks.
Early LLMs struggled to hold a plan in mind over many steps, couldn't recover from mistakes, and were limited in their ability to use external tools.
These developments enabled AI systems to work autonomously at solving real problems.
METR, an independent research group, measures how long an AI system can work autonomously on a software task using human time as a benchmark.
The latest points on the curve indicate a doubling every two months for task handling.
Agents were unreliable until November last year (2025).
Tool use allowed models to write code, search the web, read/write files, and control software.
Early LLMs could write fluently and answer questions, making them effective as chatbots.
Tool use was the second big breakthrough, giving models the ability to interact with the digital world.
The analytics platform CEO estimates 50 percent productivity gains in R&D and 20 percent in sales due to AI agents.
Capital expenditure from Amazon, Alphabet, Meta, Microsoft, and Oracle is expected to be more than $750bn this calendar year (2026).
Reasoning models became better at solving hard problems, particularly in areas requiring complex logic like maths and coding.
If the analytics company fully deployed coding agents, consumption would increase by 25-50 times.
Another AI-related company CEO mentioned a similar timeline, reaching a tipping point in agent capabilities in coding late last year (2025).
The same model that can debug code can also read financial filings, build models, flag legal risks, or plan logistics routes.
From Opus 4.5 onward, the length of tasks models could handle leapt off the trendline.
From 2019 to 2024, the length of tasks models could handle roughly doubled every seven months.
The acceleration rate for task handling then increased to doubling every three months.
OpenClaw has huge security risks due to unfettered access to digital life.
Today's leading-edge models can complete tasks that would take a skilled human roughly 12 hours, succeeding about half the time.
Three years ago, that figure (for 50% success rate) was closer to four minutes.
At 80 percent reliability, models have moved from handling tasks of about one minute to over an hour of human work.
ChatGPT was a breakthrough that led to early LLMs.
Claude Opus 4.5 could sustain autonomous coding sessions for 30 minutes-plus, navigating complex multi-file codebases, debugging errors, and producing production-ready code.
The ROI on AI tokens exceeds the return from spending the same amount on employee pay.
Previously, an application like Claude Code on a coding task was hit-or-miss; now it is mostly hit.
Reasoning models could tackle problems methodically, step-by-step.
A CEO at a portfolio company noted engineers are spending more on AI tokens than their annual compensation.
Anthropic is the fastest growing software company in history.
The entire SaaS industry tends to add about $2bn in net new annual recurring revenues (ARR) each quarter.
A single agent session can burn through more compute than a thousand chatbot queries.
This capital expenditure is up more than 50 percent year-over-year.
About three quarters of this capital expenditure is going on compute.
The Apollo program cost more than $300bn in today's money, spread over more than a decade.
All hyperscalers report that customers are absorbing AI capacity as fast as it can be deployed.
Claude Opus 4.5's benchmarked performance was good, but standard benchmarks failed to capture its real-world performance across long autonomous workflows.
The recent demand spike driven by agentic AI has mostly been in software development.
The capabilities that made coding agents possible are general and will likely spread to other domains.
Agents work differently, taking a goal and entering a loop that consumes hundreds of thousands of tokens per session.
Each new domain requires fine-tuning and new tools, but both are being built at speed.
In January (2026), an AI agent (OpenClaw) bought a car for a software engineer, negotiating a price thousands of dollars less than list.
OpenClaw is a locally-hosted AI model that connects to messaging apps and can autonomously manage email, calendars, files, and commands.
OpenClaw is an open-source project, formerly Clawdbot, that went viral earlier this year (2026).
A social network for agents called Moltbook formed, with almost a million autonomous agents interacting.
In the old paradigm, a user inputs text, and the model generates a response, consuming a few hundred or thousand tokens.
Anthropic released Claude Opus 4.5 in November last year (2025), which brought coding capabilities together at a surprising level.
Chain-of-thought reasoning, launched with OpenAI's o1 model, was one innovation that changed LLM capabilities.
The demand for a personal assistant that can act on your behalf and save you money and time is enormous.
OpenAI hired Peter Steinberger, the solo developer behind OpenClaw.
Anthropic is working to ship components like dispatching tasks from a phone to a computer and model control of operating systems/apps.
AI is much more transformational than cloud computing, which was constrained by existing IT workloads.
LLMs develop internal representations of language, logic, and knowledge that are surprisingly general.