Recently, the AI world has been a bit boring.

Back when GPT-3.5 upgraded to GPT-4, it stunned everyone. Later, multimodal models started drawing images that were indistinguishable from reality, each time causing a wave of amazement.

But now, models are getting smarter, yet the surprises are fewer and fewer.

For tasks like writing code or looking up information, today's models are already quite capable. Even if benchmark scores rise a few more points, it's hard for ordinary users to experience the kind of qualitative change they felt back then.

So it's easy to get this feeling: what else can model upgrades even bring? Are large models reaching their limit?

The laziest way to judge this is to watch the most cutting-edge companies and see where they're putting their resources.

Let's start with the currently hot Anthropic.

What Anthropic Thinks Its Selling Point Is

In June 2026, Anthropic released Claude Fable 5.

I originally thought it would, like past large model releases, put benchmarks in the most prominent position, telling users how many points the new model had improved.

But the Fable 5 official site first emphasizes not its scores, but three usage scenarios:

Agents, Coding, Enterprise workflows.

Previously, the selling point of large models was "what questions it can answer," but now Anthropic wants to prove: how big a task you can hand over to it. What Anthropic wants to sell has fundamentally changed.

Agents emphasize whether the model can autonomously advance tasks over long periods. Anthropic's new selling point is that Fable 5 can work continuously for several days, plan different stages on its own, delegate tasks to sub-agents, and check its own work.

Coding emphasizes more than just generating code. The selling point written on Anthropic's site is that Fable 5 can handle large-scale migrations, complex implementations, and multi-day autonomous programming tasks, and can also write its own tests, check results, and validate output against the final goal.

Enterprise workflows emphasize complex, multi-stage knowledge work. The model can complete deep research, analysis, and final deliverables with very little supervision. Teams just need to hand over large projects and come back for final acceptance, rather than supervising every step.

So, Agents, Coding, and Enterprise workflows are not three isolated selling points.

They actually represent a shift in the selling point:

Claude will transform from a model that answers questions into a system that can directly take over tasks.

Why "Working Independently Longer" Is So Hard

Models are clearly very smart, so why can't we just hand work over to them? This is actually a simple probability calculation:

Assume a model has a 95% probability of getting each step right, which already seems very reliable. But after executing 50 consecutive steps, the probability of getting everything right is less than 8%.

Even if the model is very powerful, with a single-step accuracy of 99%, after 100 consecutive steps, the overall success rate is only 37%.

This is, of course, a simplified calculation. In reality, errors are not completely independent, and some errors can be discovered and fixed in time. But it's enough to illustrate the fundamental difficulty of long-range agents:

Doing well on every single step does not mean the whole thing can be done well.

A Roadmap Hidden on the Official Site

In June 2026, Anthropic released a report called the Anthropic Economic Index: Cadences.

This is not an ordinary industry white paper; it's Anthropic's official research project studying how AI enters real work, professions, and economic activities. Its greatest value is that it possesses first-hand data that is almost impossible for outsiders to obtain.

This report points out a change right at the beginning:

A year ago, the main way Claude was used was back-and-forth conversation between user and assistant; but with the rapid growth of Claude Code, more and more Claude usage has become long-running Agent tasks, and traditional conversation is no longer even sufficient to describe how people are using AI.

The report contains a very intuitive set of data.

For writing a blog post or article, in a normal Chat, the median number of interactions between user and Claude is 13 rounds; in Claude Code, over half of tasks require only 1 human instruction. The subsequent breakdown, execution, and advancement are mainly completed by Claude itself.

This means that the way Claude is used has begun to shift from "human directing step-by-step" to "human states the goal, AI advances on its own."

And user expectations for this change are running even further ahead of reality.

In a survey of about 9,700 Claude users, nearly 60% believe that in the next 12 months, the proportion of work AI can complete independently will continue to rise; over one-third believe that by next year, AI will be able to complete most or even almost all of their work tasks.

This shows that what users want is no longer a chat companion, but to hand off work.

And the basic unit of Claude usage is shifting from a conversation to a delegation.

The next generation of Claude that Anthropic is betting on is not one that answers questions better, but one that can finish an entire task.

The Next Step Is Not Chatting, but Taking Over Work

The previous report shows that users are shifting from "chatting with AI" to "handing tasks to AI."

Anthropic's own product layout is also unfolding entirely in this direction.

Claude Code is the most familiar example. It allows users to hand an entire development task directly to the model, rather than just having the AI write a few lines of code.

Subsequently launched, Cowork extended this model from programmers to white-collar workers. Users can directly hand Claude a batch of files, a research task, or a task that requires cross-tool completion, and it will break it down, execute, and organize the results on its own.

At the enterprise level, Anthropic's intentions are even clearer. It has started connecting Claude to enterprise tools like Google Drive, Gmail, DocuSign, FactSet, and provides plugins for scenarios in finance, engineering, and human resources. The direction is not to make a smarter office assistant, but to let Claude truly enter a company's business processes.

Piecing these actions together, the roadmap is already very clear:

Claude Code first takes over an entire segment of development work, Cowork then extends this capability to white-collar work, and finally, through enterprise tools and workflows, it enters the company's real business.

What Anthropic wants to do is no longer a better chatbot, but a work system that can be directly dispatched by enterprises.

And Anthropic is not the only one betting on this path.

OpenAI's recently launched ChatGPT Work is also merging ChatGPT and Codex into a single work portal, allowing ordinary users to directly generate documents, websites, and presentations, rather than just asking questions in a chat box.

Microsoft's moves are even heavier. It has specifically assembled a team of about 6,000 people, sending engineers and consultants into enterprises to help clients integrate AI into actual processes like sales, customer service, finance, and R&D.

This shows that what the big players are competing for now is no longer whose chat box answers more beautifully.

The real battlefield has become who can get AI into the enterprise and take over an entire segment of work.

Large Models Have Not Reached Their End

Returning to the beginning, the progress of the next generation of large models may no longer be like the upgrade from GPT-3.5 to GPT-4, where ordinary users can see at a glance that "it suddenly got smarter."

The changes will be hidden in less conspicuous places:

Previously it could only work independently for ten minutes, now it can work for an hour; previously it needed constant human correction, now it only needs a final check; previously it could only complete a clearly defined small task, now it can handle large projects with vague goals and complex processes; previously it had to start over after a failure, now it can discover its own errors, roll back, and continue.

It doesn't seem to have suddenly become more eloquent, but it is transforming into a production capability that can be called upon, managed, and inspected.

So, if you only look at the chat box, large models have indeed entered diminishing marginal returns. But connecting the actions of Anthropic, OpenAI, and Microsoft, large models are far from reaching their end.

It's just that the scale of competition has changed.

In the past, the competition was about who answered more questions correctly; next, it will be about who can work independently for longer.

In the past, the value of a model was to give a person an answer; next, the value of a model is to take on a complete segment of work for a person.

The endgame for large models is not to become a smarter search box, but to become a new labor force.