Stanford just published the numbers I’ve been feeling in my hands for months.

As of March 2026, the top four models sit within 25 Elo points of each other. Anthropic at 1,503, xAI at 1,495, Google at 1,494, OpenAI at 1,481. The top US model leads the top Chinese model by 2.7%. Closed leads open by 3.3%. That’s the entire spread at the frontier, according to the AI Index 2026.

Stanford’s own read: capability has stopped being a clear differentiator. Competition is shifting to cost, reliability and real-world usefulness.

I read that and felt something rare… agreement between a benchmark and my daily experience.

What I Feel Every Day, Now Confirmed by Data

I run an AI consultancy. I’m in the models every single day, across software engineering, data pipelines, document processing, operations work for clients whose jobs depend on the outcome. And for a while now, I’ve stopped caring which model I’m using.

Chinese models, American models, open, closed… at the intelligence level, the difference has become marginal for real work. You can pick any of them and get comparable raw capability. You can combine them and cover the gaps. That holds for software engineering, data engineering, product management, marketing, sales, finance, logistics. Pick a function and the intelligence is already there.

I used to plan projects around model releases. I’d think, once the next generation lands, this workflow becomes feasible. I’ve stopped doing that, because after testing the same tasks across providers, I realised the ceiling I was hitting was never the model.

Waiting for a better model is no longer a strategy. It’s a delay tactic dressed up as prudence.

The models are good enough. The engineering around them is where the work remains.

The Harness Is the Product Now

Here’s what everyone gets wrong about this moment. Teams keep evaluating models like they’re choosing a supplier, running bake-offs, building comparison matrices, debating 2.7% differences in raw capability. Meanwhile the thing that actually determines whether their project ships sits completely unexamined.

That thing is the harness. The system wrapped around the model. Retrieval, tool access, memory, evaluation, failure handling, the loop that takes raw intelligence and turns it into a reliable, repeatable outcome.

The framing I keep coming back to is simple: agent equals model plus harness. The model supplies raw intelligence. The harness determines real-world usefulness.

The data backs this up in a way that should make every decision-maker pause. Only 11% of use cases reach production. The models perform fine in those failed projects. The orchestration layer collapses under production load… failure cascades, cost runaway, context overflow, version conflicts. The gap is engineering robustness, and it stays invisible until you’re three months in and the pilot quietly dies.

I’ve watched this happen at client after client. A specific one comes to mind: a mid-sized firm, sharp department head, real budget. They’d spent months comparing providers before I arrived. Two weeks of harness work later, the “worse” model they’d rejected was outperforming their shortlisted favourite inside a properly engineered loop. Most of that value came from failure handling and evaluation pipelines… boring work nobody posts about.

The loop is the leverage. Build the loop, test the loop, improve the loop. That’s the frontier now, and it moves at the speed of your engineering, at the same model price everyone else pays.

Tactical Programming Is Solved. And that Is a Relief.

The second thing the report confirmed hit closer to my identity as an engineer.

On SWE-bench Verified, the benchmark for autonomous software engineering, performance rose from roughly 60% in 2024 to close to 100% in 2025. Read that again. The benchmark that separated frontier systems a year ago can no longer tell them apart. Tactical coding, the mechanical act of producing correct syntax, is functionally solved.

If your satisfaction came from writing code by hand, from memorising the syntax of a particular language, that problem has been handed to the machines.

Matt Pocock put it in words that resonated with me deeply:

“Tactical programming is dead. Strategic programming has never been more vital. If you’re lamenting writing code by hand like something you love is being taken away from you, see if you can fall in love with solving problems instead. Maybe you’ll find out that’s what you really loved the whole time.”

I read that and recognised my entire career. I always loved the problem more than the syntax. The code was a means to an end… a way to make a document pipeline stop haemorrhaging money, a way to turn a chaotic, messy, unmanageable compliance process into something a business could actually rely on. Syntax was the toll I paid to get there.

The toll booth just closed. The destination still needs reaching.

And the destination got harder, in the best possible way. Around 90% of code is now AI-generated, and that code still requires human architectural guidance and review. The bottleneck moved from typing to judgement. From syntax to system design. From “can you write this function” to “do you understand how this system should fail, degrade, recover and scale”.

The Skill That Survives

So where does this leave you, if your work touches technology at all… and in 2026, it does.

It depends on which layer of the work you occupy. If your role was the mechanical version of a function, translating clear requirements into standard code, filling templates, moving data by hand, that layer is exposed. The AI Index makes that impossible to ignore.

The durable skill that survives is orchestration. Conducting several AI systems towards an outcome. Owning the loop end to end. Thinking in failure modes, cost curves, evaluation criteria and architectural trade-offs while the models handle the tactical execution underneath you.

In gaming terms, the tutorial phase of AI adoption is over and everyone got the same power-ups. The differentiation now comes from how you build the skill tree around them.

Here’s what that looks like in practice:

• Stop benchmarking models and start benchmarking your loops. Measure how your system handles failure, cost and drift under real load.

• Invest in evaluation before you invest in scale. A harness you can’t measure is a harness you can’t trust.

• Treat model choice as a swappable component. The 25-point Elo spread means your architecture should survive changing providers in an afternoon.

• Redirect your team’s energy from writing code to designing systems. The judgement layer is where your people compound in value.

I highly recommend you run one honest audit this month: take your most important AI initiative and ask whether the risk sits in the model or in the engineering around it. In every audit I’ve run this year, the answer sat in the engineering.

You Will Never Run Out of Problems

There’s a quiet anxiety underneath all of this. I hear it in conversations with engineers and executives alike… the fear that once the machines write the code, the meaningful work disappears.

My experience points the other way, and the economics agree. The estimated value of generative AI to US consumers reached 172 billion dollars annually by early 2026, with the median value per user tripling in a single year. That value flows to the people who moved past waiting and started engineering.

Problem-solving has no benchmark that saturates. There’s no SWE-bench for “understand this specific business, its constraints, its people, and design the system that transforms it”. That work compounds, and it never finishes.

You will never run out of ideas. You will never run out of problems worth solving. The models cleared the tactical layer so you could finally spend your full attention on the strategic one… and it’s going to be hard, but you’ll make it work, because the hard part was always the part worth doing.

The frontier converged. The loop is open.

So here’s my question for you: when you look at your own work honestly, what did you actually love… the syntax, or the solving?