OpenAI and Anthropic Agree: In 2026, 'Capability Overhang' Matters More Than 'Better Models'

Last week while scrolling through Twitter, I came across an interesting tweet from OpenAI.

They introduced a concept called ‘capability overhang.’ Simply put: the gap between what today’s AI models can do and what most people actually use them for is massive.

What’s fascinating is that in the same week, Jack Clark, co-founder of Anthropic, wrote almost the same thing in his Import AI newsletter.

The makers of ChatGPT and Claude reached the same conclusion at the same time.

That alone is worth thinking about.

What is ‘Capability Overhang’?

Here’s an example. GPT-4 level models can theoretically handle complex tasks: analyzing lengthy documents, writing code, data analysis, even drafting business plans.

But in reality, what do most people use them for?

Writing emails, polishing resumes, asking simple questions.

Not that these uses are bad—but the ceiling of these models far exceeds most people’s usage habits.

OpenAI’s exact words: ‘In 2026, AGI progress isn’t just about how powerful models are, but whether ordinary people can actually use them well.’

That’s refreshingly honest.

Why Are Both Saying This Now?

I can think of a few reasons.

First, model iteration is slowing down. Not that progress has stopped, but marginal returns are diminishing. The jump from GPT-4 to GPT-5 isn’t as dramatic as GPT-3 to GPT-4.

Second, the competitive landscape is shifting. Anthropic’s Claude is doing well in enterprise markets, Google’s Gemini is catching up. As model capabilities converge, differentiation shifts from ‘who’s smarter’ to ‘who’s more usable.’

Third, and most practically—commercial pressure. Training large models is incredibly expensive. If users only write weekly reports with them, the ROI doesn’t work. Users must tap into deeper capabilities to justify the cost.

What Does This Mean for Regular People?

Honestly, I think this is good news.

Over the past year, many people felt anxious: ‘AI is developing so fast, will I be replaced?’

But ‘capability overhang’ shows that the problem isn’t AI isn’t strong enough—it’s that most people don’t know how to use it yet.

In other words, mastering how to use AI is itself a competitive advantage.

The same model can produce vastly different outputs in different hands.

How to Unlock Deeper Model Capabilities

Let me share some practices that have worked for me.

First, shift from ‘one-shot’ to ‘multi-turn collaboration.’

Don’t expect perfect answers from a single prompt. Treat AI like an intern—give it context, provide feedback on intermediate results, refine iteratively.

Second, try complex tasks.

Many people only use AI for simple things like ‘help me fix this headline.’ Try something complex: ‘Analyze the three core conclusions of this report and identify weak points in the data support.’

Third, build a personal prompt library.

Good prompts are reusable. Organize common scenarios into templates for future use.

Final Thoughts

OpenAI and Anthropic reaching this consensus marks a turning point for the AI industry.

Previously, everyone competed on ‘my model is stronger than yours.’ Now it’s becoming ‘my users are better at using it than yours.’

This shift is actually better for users.

After all, we don’t need to know how to build the engine—just how to drive the car well.