In February last year, I was using AI coding tools inside VS Code. They could help with an individual file, but changes that depended on several parts of a repository were much harder to get right. A year later, the difference in what I can ask them to do feels enormous.

That is my experience of working with the tools, rather than a controlled comparison of models. But it has changed how I think about the gap between having access to AI and getting useful work out of it.

On 5 February, OpenAI launched Frontier, a platform for building, deploying and managing AI agents inside enterprises. The same day, Anthropic released Claude Opus 4.6, with improvements in coding, long-running agentic tasks and knowledge work.

The obvious interpretation is that businesses now have access to better AI. I think the more interesting interpretation is that access is becoming less important as a differentiator.

If two companies can buy broadly the same intelligence, the gap between them increasingly comes down to what they do with it. That is where I think the next competitive divide is opening.

Access is no longer the problem

For much of the history of enterprise technology, capability was scarce. Sophisticated systems required specialist teams, large budgets, long implementation programmes and significant infrastructure. Smaller businesses could often see what larger competitors were doing without being able to reproduce it.

Frontier AI changes some of those economics. A company does not need to train its own frontier model to benefit from one. It can access OpenAI, Anthropic and others through products and APIs. Improvements made by the labs can reach millions of users quickly.

In February, OpenAI described Frontier as infrastructure for AI coworkers that can operate with shared context, permissions and boundaries. Anthropic’s Opus 4.6 pushed further into sustained agentic work.

These are not identical products. But they point in the same direction. Powerful intelligence is becoming a purchasable input. That sounds like an equaliser. It may not be.

Equal tools produce unequal outcomes

An enterprise contact recently told me how pleased they were with Copilot for rewriting emails. That is a useful application. It also reminded me how different our experiences of the same technology can be. I am connecting models to GitHub, deployment tools and other systems so they can help carry work through. We can both say we use AI, while meaning quite different things.

I also have the time and freedom to experiment. An employee in a large company may have neither. They work with centrally approved tools, data policies and processes they cannot change on their own. It would be unfair to mistake those constraints for a lack of imagination.

Give two businesses the same CRM and they will not become equally good at sales. Give two factories the same machinery and they will not produce identical economics. Give two people the same spreadsheet and one may build a useful financial model while the other creates a mess. Technology matters. The system around it matters more than we sometimes admit.

AI amplifies this because its possible applications are so broad. One organisation can use a frontier model as a slightly better writing assistant. Another can connect it to data, tools and processes and delegate meaningful parts of a workflow. Both can truthfully say they are “using AI”. Economically, they may be doing completely different things.

OpenAI’s own launch material made the point unusually clearly: the constraint it sees in leading enterprises is increasingly not model intelligence but how agents are built and run inside organisations. That is a significant shift.

The AI gap is moving from who has access to intelligence towards who can reorganise work around it.

The unit of value is changing

A lot of early enterprise AI adoption was measured in seats. How many employees have access? How many people are active? How many prompts are being sent? Those numbers are useful for understanding adoption. They are much less useful for understanding value. I would rather know what changed.

  • Did a process that took three days take three hours?
  • Did salespeople spend more time with customers?
  • Did a team increase throughput without increasing headcount?
  • Did a previously uneconomic service become viable?
  • Did the quality of a decision improve?

Those questions force the conversation away from novelty and towards economics.

OpenAI gave examples in its Frontier announcement of agents being used for production optimisation, sales processes and energy output. The exact results belong to those organisations, not every company that installs an agent platform.

But the pattern matters. The ambition is no longer merely to help an employee write the email faster. It is to change how work gets done.

This is an organisational capability

That creates a problem for businesses looking for an “AI strategy”. Buying the technology is comparatively easy.

  • Finding valuable problems is harder.
  • Redesigning a workflow is harder.
  • Connecting the right systems and data is harder.
  • Deciding where a human should remain accountable is harder.
  • Measuring whether the change created value is harder.
  • And getting people to adopt a different way of working can be harder than all of them.

Later in February, OpenAI announced Frontier Alliances with consultancies including BCG, McKinsey, Accenture and Capgemini. The announcement explicitly highlighted leadership alignment, workflow redesign, systems integration and change management.

That makes sense to me. Connecting applications to a model has changed what I can do more than simply adding another AI feature inside an application. Inside a business, making those connections involves people who understand the work as well as the technology. Buying access is only the beginning.

Fast followers may not stay close

There is a comforting assumption in technology adoption that being a fast follower is enough. Sometimes it is. But AI may have a compounding characteristic that makes delay more expensive.

An organisation that deploys useful agents does not only receive the immediate productivity gain. It also learns.

  • It learns which workflows are suitable.
  • It learns where models fail.
  • It learns how to structure context.
  • It learns what governance is actually required rather than theoretically required.
  • It learns how employees respond.
  • It develops reusable integrations, evaluation methods and implementation muscle.

The next project can therefore become easier than the first. Meanwhile, the organisation still running disconnected pilots is learning at a slower rate. That is how a capability gap can become an opportunity gap.

The model race can be a distraction

OpenAI and Anthropic will continue competing. Google and others will too. There will be benchmark charts, price changes, context windows and new model releases. Some of those differences will matter enormously for particular tasks.

But a CEO can spend an extraordinary amount of time following the model race without changing a single important process inside the business. That is why I think model selection is becoming a smaller part of the strategic question.

A sufficiently capable model with the right context, tools, permissions, evaluation and workflow may create far more value than the nominally “best” model dropped into a weak operating environment. The technology is improving too quickly to build strategy around a permanent winner anyway. The more durable capability is knowing how to turn intelligence into an outcome.

Start with the gap that matters

The phrase “AI opportunity gap” can sound like another way of saying some companies are ahead. I think it is more specific than that. The gap is between available capability and captured value.

That gap exists inside individual companies as much as between them. A business may already pay for extraordinarily capable AI while using only a fraction of what it could economically exploit.

The practical question is therefore not “Are we using the latest model?” It is: Where could the intelligence already available to us materially change the economics of this business?

Then find one meaningful workflow. Understand it properly. Redesign it. Measure the result. Learn from the failure modes. Repeat.

Frontier models will keep moving. The companies that learn how to convert that movement into value may move faster still.

Sources: OpenAI Frontier · Claude Opus 4.6 · OpenAI Frontier Alliances