AI Got Cheaper and Smarter for 30 Months — Your 2027 Plan

What Actually Changed in the Last 30 Months?

For a stretch in 2023 and early 2024, one model was simply the best: GPT-4 had no serious rival. That is over. Mollick counts roughly 20 different models that have held the top spot at one point or another since then, all from a small set of labs. Below the frontier, about 180 models are now better than GPT-4 was at its release, many of them open-weight, and around 19 are small enough to run on a modern smartphone.

That churn matters more for planning than any single model's launch. It means the “best AI” is a moving target that changes hands every few months, and that near-frontier capability has become cheap and widely available rather than locked behind one vendor. If you want to watch this happen in near real time, independent leaderboards like Artificial Analysis track hundreds of models across intelligence, speed, and cost per task and update constantly — the ranking you bookmark in January will look different by spring.

The honest caveat, which Mollick makes himself: progress is “improving exponentially but unevenly.” Math moved from terrible to superhuman in the same window that long-form writing stayed mediocre. The frontier is jagged. A model that can do an expert task flawlessly may fumble something a child handles easily, and those small gaps still create large, ongoing needs for human involvement. That jaggedness is the whole reason a thoughtful deployment beats a blind one — and it is why turning AI pilots into AI Employees is an execution problem, not a procurement one.

How Much Cheaper and More Capable Did AI Actually Get?

Two curves carry most of the story: how long a task AI can finish on its own, and how little it costs to get a correct answer. Both are measured by independent research groups, not by the labs selling the models.

On task length, the benchmark to know is METR's long-horizon measure, which tracks the length of human work an AI can complete with at least a 50% chance of success from a single prompt. METR's original 2025 analysis found this horizon had been doubling roughly every seven months for years. Mollick's charts carry it forward: the horizon doubled about eight times over the measured period, from a task that takes a person about four minutes in 2024 to one taking about 17 hours when he submitted his draft in April 2026. METR has since said it can no longer measure the figure reliably with its current task suite — the models have outrun the ruler. Its updated Time Horizon tracking still puts the best model's 50% horizon in the 11-to-16-hour range as of mid-2026, while cautioning that anything above 16 hours is unreliable to measure.

On cost, the picture is just as steep. Mollick cites a GPQA “Google-Proof” science exam example in which a single correct answer that cost about 30 cents in early 2025 later cost roughly four hundredths of a cent — a drop of about 750x. That specific example traces to Epoch AI's cost research; TechSpot's coverage of Epoch's pricing analysis reports that, at a fixed performance level, model costs have fallen roughly 47% per quarter — about 13x per year — since 2023, with the GPQA example running from a January 2025 model at $0.30 per question to a 2026 model matching the score at $0.0004. (Treat the 750x as an approximate ratio derived from those two endpoints, not a precise constant.) Epoch's own inference-price work found the price to hit a given capability milestone has dropped anywhere from 9x to 900x per year depending on the task.

Here is the trajectory in one view:

Two caveats before you lean on any of these. First, the cost figure measures cost at a fixed performance level — a newer model is not automatically cheaper for everything, only for matching an older result. Second, the fastest price drops happened recently and may not persist at the same rate; TechSpot's account notes the per-quarter decline has already halved for mature benchmarks. Plan for the direction, not the exact multiple.

Why Does the Chatbot-to-Agent Shift Change Your 2027 Plan?

The capability curves describe what the models can do. The more important planning signal is how people use them. Mollick frames the shift as a move “from chatbots to agents that do real work,” and he adds a third working mode to the “centaur” and “cyborg” patterns from his earlier writing: the “conjurer” — someone who, in his words, “summon[s] agents when needed, assign[s] them work, and evaluate[s] what comes back.” He calls this “the most powerful way to use AI.”

For a business, that is the difference between buying a smarter search box and standing up a worker. A chatbot answers a question; an agent runs a multi-step process — pulling data, drafting the output, taking the next action — and reports back. When the frontier task horizon was four minutes, “agent” was mostly a demo. At multi-hour horizons it is a staffing decision, which is exactly why your 2027 headcount plan should model AI Employees as real capacity rather than as a software line item.

This is also where the curves stop being abstract and start being budgetary. If per-task cost is falling ~13x a year, the right planning assumption is that the agentic workflow you pilot in early 2027 will be meaningfully cheaper to run by year-end — so the question is not “can we afford to run this at scale?” but “are we ready to supervise it at scale?” As we argued in our look at why waiting is the real risk, the cost of sitting out is rarely the technology price — it is the operating experience you do not build while competitors do.

What Stays Human — and Why That Is Your Governance Story?

Mollick is careful not to oversell. He names four durable human advantages — “deep knowledge, wide knowledge, taste, and agency” — and his new “conjurer” mode comes with an explicit warning: as agents take on more, “we lose control over what, exactly, the AI is doing.” That is not a reason to avoid agents. It is the design brief for deploying them responsibly.

Falling cost and rising task length make it trivial to point an agent at a real workflow. What does not get cheaper is the oversight — knowing which outputs to trust, where the agent has authority to act, and what it is doing with your data while it acts. The more work you hand off, the more that supervision becomes the job. We have written about how the agents you deploy become insiders with real access and real reach, and the point holds here: the human edge Mollick describes is only an edge if you build the governance to exercise it.

Practically, that means a secure gateway between your agents and your systems, clear boundaries on what each agent may touch, and measurement that tells you whether the work is actually good — not just whether it got done. Capability you cannot supervise is a liability, not an asset. The planning takeaway is that your 2027 AI budget needs two lines, not one: the falling cost of running agents, and the rising investment in governing them.

What Should a Fort Wayne or Northeast Indiana Business Do With These Curves in 2027?

For a professional-services firm, manufacturer, or home-services operator in Allen County, DeKalb County, or the wider Fort Wayne region, the charts translate into three concrete moves — and none of them require being first to the frontier.

Re-review your model and vendor choice quarterly, not annually. With about 20 models having traded the #1 spot and roughly 180 now beating the old GPT-4, the “best” tool for your use case will change inside a budget year. Lock in flexibility, not a single vendor, and schedule the review.

Budget for falling per-task cost, not rising license fees. If running an agentic workflow is on a ~13x-per-year cost curve, a pilot that looks marginal on today's prices may be clearly worth it by mid-2027. Price the trajectory into the plan rather than the spot cost.

Pilot one agentic workflow now instead of waiting for the frontier to settle. It will not settle on your timeline. The operators who will be fluent with AI Employees in 2027 are the ones running a governed, measured pilot in the first quarter — which is precisely what the practical adoption playbook for Fort Wayne owners is built to help you do. Start with one bounded process — lead intake, document review, after-hours phone coverage — give the agent clear limits, and measure it against the work it replaces.

The firms that benefit most here are not the biggest. They are the ones that treat a moving curve as a reason to plan in quarters, build the governance early, and let falling costs do the heavy lifting on ROI.

Build Your 2027 Plan Around the Curve, Not Today's Prices

The trajectory is clear even where the exact multiples are fuzzy: AI is getting more capable and cheaper fast enough that a static 2027 plan is a liability. The businesses that win the next year are the ones that pilot a governed agentic workflow now, measure it honestly, and revise on a quarterly cadence.

That is the entire premise of Cloud Radix AI Employees — autonomous agents deployed behind a secure gateway, pointed at one real workflow, and measured against the work they take over. If you are a Fort Wayne or Northeast Indiana business turning these curves into a 2027 operating plan, we will help you choose the first workflow, set the guardrails, and track whether it actually pays. Reach out and we will map your first pilot — and, just as important, the oversight that keeps it safe.

Frequently Asked Questions

Sources & Further Reading

• One Useful Thing (Ethan Mollick): oneusefulthing.org/p/30-months-on-the-exponential-in-charts — 30 Months on the Exponential, in Charts.
• METR: metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks — Measuring AI Ability to Complete Long Tasks.
• Epoch AI: epoch.ai/blog/llm-inference-price-trends — LLM Inference Prices Have Fallen Rapidly but Unequally Across Tasks.
• TechSpot: techspot.com/news/114055-cost-ai-performance-falling-fast — AI Is Getting Cheaper Faster Than Almost Any Technology in History.
• Artificial Analysis: artificialanalysis.ai — Independent Analysis of AI Models.
• Wikipedia: en.wikipedia.org/wiki/METR — METR (Time Horizon benchmark overview).