Chatbot Twilight: Opus 4.7 Does a Programmer's 2 Weeks of Work for $251

On June 30, 2026, Wharton professor Ethan Mollick published an analysis that marks a turning point: the era of chatbots is ending, replaced by autonomous AI agents that can work for hours without human involvement. This is already happening inside the largest AI labs.

What agents can do today

In an Epoch AI study, Anthropic's Opus 4.7 worked on its own for 14 hours and produced software that would have taken humans 2 to 17 weeks to develop. The cost was $251 in tokens.

• Epoch AI (2026): Opus 4.7 worked autonomously for 14 hours, the equivalent of 2–17 weeks of engineering labor, for $251
• Mollick's experiments: Claude Fable completed complex technical projects in 9 hours that had taken a team a week
• METR and the UK's AI Security Institute record AI autonomy growing faster than any forecast
• OpenAI: a quarter of the company's employees manage at least 4 agents every week
• The continuous-work horizon for AI grew from 2 hours to 16+ hours per prompt in a matter of months

Mollick stresses that AI systems do not yet pass every test, "but they are unambiguously improving at a very high rate." Several independent benchmarks confirm this, including GDPval, which compares AI performance with that of human experts across different industries.

Why experts matter more than programmers

An internal OpenAI study found an unexpected pattern: success with Claude Code did not depend on professional background. Lawyers did as well as developers.

"The more domain experience a person has, the more successfully they use Claude Code in their field.

And, more interesting still, the more useful output they get from each prompt," says Ethan Mollick's analysis, citing OpenAI data.

When writing code becomes the agent's job, professional boundaries blur: a lawyer, doctor or financial analyst gets the same tools as a developer. The advantage now belongs to people who understand their field deeply, not to people who know how to code. To test qualitative differences between models, Mollick used an unconventional test: he asked AI to build interactive simulations of a seaport. The models diverged substantially in design and judgment, qualities that standard benchmarks do not capture.

Two tiers of models have formed in parallel. American closed models (Anthropic, OpenAI, Google) sit at the top of the performance range. Chinese open-weights models trail by 6–12 months, but they are significantly cheaper and can be deployed locally.

Why companies can't keep up with the change

Institutions move at human speed: approvals, committees, quarters. AI improves along an inhuman exponential, and the gap widens with every cycle.

"Instability is what happens when institutions moving at the speed of people (or, worse, committees) try to track a capability curve that is inherently completely non-human," —

Ethan Mollick, One Useful Thing.

At the start of 2025, an AI agent could work autonomously for about 2 hours, with a high error rate. By mid-2026, that figure was 16+ hours per prompt. AI strategies written in late 2024 now describe an obsolete reality. Concrete examples include emergency regulatory measures in cybersecurity after a sudden jump in AI capabilities, and sharp stock-market swings when AI suddenly threatened the business models of entire sectors.

What this means

We are not entering a transition period before things stabilize; we are entering a permanent state of acceleration. The winners will be those who can set tasks for agents, evaluate their results, and know their professional field deeply.

Frequently asked questions

How much does it cost to run an AI agent on a large project?

According to Epoch AI, Opus 4.7 completed the equivalent of 2–17 weeks of engineering work for $251 in tokens. The final cost depends on the model, the duration of the task and the number of prompts.

How does an AI agent differ from a chatbot?

A chatbot answers one question and waits for the next. An agent receives a task, breaks it into steps on its own, corrects errors along the way and works for hours without user intervention. This is how Anthropic's Claude Code and OpenAI's Codex work.