By late 2024, you could rent the kind of language model that once cost tens of millions of dollars to train for a few cents per query. Stanford’s AI Index measured the drop at roughly 280-fold in under two years. That single number explains more about where artificial intelligence is heading than any roadmap a lab has published. The interesting question is no longer whether the models are impressive. It is what happens when something that capable becomes cheap, fast, and quietly embedded in the software people already use.
Chatbots were the warm-up. Agents are the shift.
A chatbot answers. An agent acts. It reads a task, breaks it into steps, opens tools, checks its own output, and keeps going until something is done. Coding agents already ship production code at companies like Anthropic, Cursor and Cognition. Support agents resolve tickets end to end. Browser agents fill in forms, compare prices, and file expenses.
The catch is arithmetic. If each step in a twenty-step task succeeds 95% of the time, the whole job lands about 36% of the time. Reliability compounds downward, which is why agent demos look like magic and agent deployments look like a project plan. Expect the next few years to be spent not on making agents smarter but on making them recover gracefully when step seven goes wrong.
Why 2025 and 2026 look different
Two things changed. Models got better at planning mid-task, and the tooling around them matured: sandboxed execution, memory stores, permission layers, audit logs. The plumbing matters more than the benchmark scores. A model that can run a shell command is useless if nobody will let it near production.
Smaller models are eating the frontier’s lunch
Every frontier release gets distilled into something a tenth the size that handles 80% of real requests. Those smaller models run on a laptop, a phone, or a single rented GPU. Open-weight families from Meta, Mistral, Alibaba and DeepSeek have made the middle of the market brutally competitive.
This is the part people miss. The future of AI is not one giant brain in a data centre answering everything. It is a tiered system: a small model on your device handling the routine, a mid-size model in the cloud for the tricky stuff, and a frontier model called in rarely, like a specialist you consult. If you want the groundwork on how these systems actually function, this plain-English guide to how artificial intelligence works is a decent starting point.
The bottleneck moved from chips to electricity
Training runs are now measured in gigawatts. The International Energy Agency has projected that global data centre electricity demand could roughly double to around 945 TWh by 2030, with AI as the main driver. Microsoft signed a deal to restart Three Mile Island. Google contracted small modular reactors. Utilities in Virginia have queues of interconnection requests stretching past the end of the decade.
That constraint reshapes strategy. Efficiency research suddenly has a business case again. So does placing inference physically close to where the power is, and so does the unglamorous work of cooling and scheduling. A lab that can serve the same model at a third of the energy cost wins on margin, not on prestige.
The AGI argument is really an argument about timelines
Ask whether machines will match humans across the board and almost everyone says yes eventually. The disagreement is about when, and that disagreement drives everything from hiring to regulation. Some researchers put it within a decade. Others point out that models still fail at tasks a five-year-old handles, particularly anything requiring physical intuition or a consistent model of the world. The race to build artificial general intelligence is worth following closely because the definition keeps shifting as capabilities arrive.
Demis Hassabis has argued that the path runs through scientific discovery rather than pure scaling, which is why DeepMind spent years on protein structure before it spent them on chat. His view, traced in this profile of Hassabis and his pursuit of AGI, treats benchmark progress as a symptom rather than the goal.
Not everyone is betting on the same architecture
Mira Murati, who led the team that built ChatGPT, left OpenAI and raised capital for Thinking Machines Lab on a thesis that current models are trained the wrong way. Her bet involves systems that adapt to individual users rather than being frozen at deployment. Read the detail in why Murati is betting against how ChatGPT works. It is a useful corrective to the assumption that the current recipe is the only one.
Meanwhile xAI, OpenAI, Google and Anthropic are locked in a spending race that looks less like research and more like infrastructure construction. The competition has personalities attached now, and they influence product decisions more than most people realise.
What this does to work, task by task
Forecasts that name a single number of jobs lost are mostly noise. The granular picture is clearer. Work is unbundling into tasks, and tasks are being reassigned.
- Junior coding changed first. Boilerplate, tests, migrations and glue code are increasingly generated and reviewed, not written. What survives is judgment about what to build.
- Customer support split into resolved-without-a-human and escalated-with-full-context. Headcount moved toward the second category.
- Legal and medical documentation absorbed the biggest time savings, because drafting was always the slow part and review was always the point.
- Physical trades stayed largely untouched. Robots are improving, but a plumber’s job is a control problem in a messy world, not a text problem.
The pattern across all of them: the value shifts to whoever can verify quality quickly. Someone still has to know when the output is wrong.
Separating the real tools from the noise
Search for any AI-adjacent problem and you will find a hundred sites claiming to solve it, most of them thin wrappers around the same API. That clutter makes evaluation a skill in itself. Useful questions: does it show its work, does it handle failure, does it cost more than the labour it replaces? There is a practical method in this piece on how to spot AI websites worth your time, and it holds up better than most vendor comparisons.
Signals worth watching over the next two years
Ignore the demo videos. Watch these instead.
- Inference price per unit of capability. If it keeps falling by roughly an order of magnitude a year, whole categories of product become viable that were not last spring.
- Agent success rates on long tasks. Not five-step tasks. Fifty-step ones. That number tells you when real automation arrives.
- Power contracts. Signed gigawatts are a better forecast of compute in 2028 than any funding round announced today.
- Court rulings on training data. A few decisions in the US and EU will determine whether model development stays cheap or becomes a licensing business.
- Whether anyone ships a reliable memory layer. Systems that remember context across months change the product category entirely.
None of these are glamorous. All of them are measurable, and each one will quietly determine what you can actually do with AI two years from now, long before the industry agrees on a definition of anything.