Сумерки чат-ботов: Opus 4.7 за $251 выполняет работу программиста на 2 недели

30 June 2026, Wharton professor Ethan Mollick published an analysis marking a turning point: the era of chatbots is ending, replaced by autonomous AI agents capable of working for hours without human involvement — and this is already happening inside the largest AI laboratories.

What agents can do today

Opus 4.7 from Anthropic, in an Epoch AI study, worked autonomously for 14 hours and produced software that would have taken humans 2 to 17 weeks to develop. The cost: $251 in tokens.

• Epoch AI (2026): Opus 4.7 worked autonomously for 14 hours — equivalent to 2–17 weeks of engineering effort for $251
• Mollick's experiments: Claude Fable closed complex technical projects in 9 hours — work that would have taken a team a full week
• METR and the AI Security Institute (UK) are recording AI autonomy growing faster than any forecast
• OpenAI: a quarter of the company's employees manage at least 4 agents every week
• The continuous working horizon of AI has grown from 2 hours to 16+ hours per single prompt in just a few months

Mollick emphasizes: AI systems do not yet pass every test, "but they are unambiguously improving at a very high rate" — this is confirmed by several independent benchmarks, including GDPval, which compares AI performance against the level of human experts across different industries.

Why domain experts matter more than programmers

An internal OpenAI study revealed an unexpected pattern: success when working with Claude Code did not depend on professional background — lawyers performed just as well as developers.

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

And — what is even more interesting — the more useful output they get from each prompt," states Ethan Mollick's analysis, citing OpenAI data.

When code-writing becomes an agent's task, professional boundaries blur: a lawyer, doctor, or financial analyst gains access to the same tools as a developer. The advantage now belongs to those who deeply understand their own field, not to those who know how to code. To test qualitative differences, Mollick used an unconventional test: he asked AI models to build interactive port simulations. The models diverged significantly in design and judgment — characteristics that standard benchmarks do not capture.

In parallel, two tiers of models have formed. American closed-source models (Anthropic, OpenAI, Google) are at peak performance. Chinese open-weights models lag 6–12 months behind, but are significantly cheaper and available for local deployment.

Why companies cannot keep up with the changes

Institutions move at the speed of people — approvals, committees, quarters. AI improves on an inhuman exponential curve, and the gap widens with each cycle.

"Instability is what happens when institutions moving at human speed (or, worse, committee speed) try to track a capability curve that is by its nature entirely non-human," —

Ethan Mollick, One Useful Thing.

In early 2025, an AI agent could work autonomously for about 2 hours with a high error rate. By mid-2026, it's already 16+ hours per single prompt. AI strategies written in late 2024 describe an outdated reality today. Concrete examples include emergency regulatory measures in cybersecurity following a sudden leap 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 transitional period before stabilization — we are entering a permanent state of acceleration. Those who win are the ones who know how to assign tasks to agents, evaluate their results, and have deep expertise in their professional field.

Frequently asked questions

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

According to Epoch AI data, Opus 4.7 completed the equivalent of 2–17 weeks of engineering work for $251 in tokens. The final cost depends on the model, task duration, and 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, independently breaks it down into steps, corrects errors along the way, and works for hours without user intervention — this is exactly how Claude Code from Anthropic and Codex from OpenAI work.