When Models Converge, the Factory Wins

A few months ago the model race felt simple in my daily work.

Anthropic had the crown. Sonnet was the best everyday model for Hermes. Opus was the model I reached for when coding needed force, taste and breakthrough energy. If I wanted a difficult agentic workflow to move, I wanted Claude in the loop.

Then Anthropic made what I still think was a strategic mistake.

It pushed the OpenClaw and Hermes power-user crowd out of its subscription model. I understand why. Compute was limited. Business customers were profitable. The company had to protect the segment that paid the bills.

But there was a hidden cost.

The builders went shopping.

These were not casual chatbot users. These were people wiring models into terminals, repos, personal operating systems, multi-agent workflows, company automation and daily execution. They were not asking the model a few clever questions. They were discovering what a frontier model becomes when it is embedded into work.

Anthropic protected its enterprise margin, but it weakened its relationship with the most important experimental frontier: the people turning models into infrastructure.

I was one of them.

I moved my Hermes agents to GPT-5.5. It was not perfect. It talked too much. It did not always have the same Claude taste. But it worked. It could chat, code, write, reason and run my life/company stack well enough that I did not need a different model for every mood.

One model for almost everything.

That mattered.

Then came GPT-5.6 Sol. Then Grok 4.5. And suddenly the whole question changed.

the convergence problem

The interesting thing is not that one model is better than another on a benchmark.

Of course there are differences. Fable 5 may still be the best raw model in certain coding and reasoning tasks. Elon Musk himself appears to treat Anthropic as the top contender. Sol feels stronger than GPT-5.5. Grok 4.5 is now shockingly usable compared to where Grok was only a few months ago.

But in my daily work, the models are converging.

Not in the abstract. In the body of work.

When I use them to build AI4 Accountancy, run Hermes, write essays, operate agents, debug workflows and structure my life, the differences are real but less decisive than they used to be. A year ago, the wrong model could break the whole experience. Today, the top models are all broadly acceptable for most serious work.

That changes the market.

When capability gaps are large, intelligence wins.

When capability gaps narrow, everything else starts to matter.

Speed matters.

Price matters.

Rate limits matter.

Token efficiency matters.

Guardrails matter.

Subscription access matters.

Latency matters when ten agents are working in parallel.

The best model is no longer simply the smartest model. It is the model you can actually use at scale without fighting the business model behind it.

This is where Anthropic's position becomes more fragile than it looks.

A brilliant model with tight limits is not a workhorse. It is a luxury instrument.

That is fine if the market stays in demo mode. It is not fine when models become operating systems.

intelligence is becoming table stakes

For a long time, the frontier model race felt like a contest of minds.

Which lab had the smartest model? Which model solved the hardest benchmark? Which one wrote the cleanest code? Which one had the best taste? Which one could handle ambiguity without collapsing into slop?

That still matters.

But the center of gravity is moving.

Once models become good enough, the question becomes industrial. Who can produce intelligence faster, cheaper, with tighter feedback loops and fewer artificial constraints?

This is why xAI is suddenly more interesting than many people expected.

A few months ago Grok did not work for my workflows. It was not a real alternative for Hermes. It felt like a loud model with potential, not a trusted daily partner.

Now Grok 4.5 is different. It is fast. It is cheaper. It is good enough that the tradeoff becomes serious. If the output is roughly comparable in daily work, and one model is faster and cheaper, the decision starts to move away from purity and toward throughput.

That is not because quality no longer matters.

It is because quality has crossed a threshold.

When every top model can reason, code, browse, write, plan and use tools well enough, the scarce resource becomes execution bandwidth.

A slightly smarter model that is expensive, limited and paternalistic can lose to a slightly less polished model that is fast, open, cheap and available.

Especially in agentic work.

One chat answer is not the unit anymore. A company is becoming a swarm of model calls. Hermes is not a chatbot. Cursor is not autocomplete. AI4 Accountancy will not be one prompt. It will be thousands of small cognitive actions stitched into workflows.

In that world, marginal cost and latency are not details. They shape what can exist.

the factory behind the mind

Elon Musk's advantage is not only that xAI has a better model today. It may or may not.

The advantage is the production philosophy.

Musk does not usually build by waiting for the perfect design and then releasing it. SpaceX did not get to Starship by treating every prototype as sacred. It built a production line, started with rough versions, blew things up, collected reality, redesigned, rebuilt and accelerated the loop.

The production line is the learning machine.

That is the signature.

Not engineering as a waterfall. Engineering as a compounding feedback loop.

That matters for AI because frontier models are not static products. They are moving industrial systems. Data, compute, inference, post-training, evals, user feedback, tool use, synthetic environments, hardware, networking, power and deployment all feed into each other.

The model is not the product.

The machine that produces the model is the product.

This is where xAI's vertical integration story becomes important. Compute first. Massive clusters. Faster deployment. Less abstraction. Fewer layers. More direct control over the stack. The ambition to own more of the production chain, maybe even down into machines, chips, factories, energy and eventually space-based infrastructure if Starship and related plans ever make that plausible.

Some of that will sound insane until it does not.

But the strategic logic is clear: reduce dependency, reduce waiting, shorten the feedback loop.

OpenAI and Anthropic are not stupid. They know they must iterate. Anthropic is arguably king right now on model quality. OpenAI has enormous distribution and remains deeply capable. Google may be the sleeping giant because it has compute, chips, infrastructure, data, research and patience.

But xAI is warming up a different kind of machine.

If frontier intelligence keeps converging, the lab that learns fastest wins.

Not the lab with the best press release.

The lab with the tightest industrial loop.

the ethics of usable intelligence

There is another layer here that is easy to dismiss as philosophical, but it becomes practical fast.

Ethics.

Not ethics as a PDF. Not the corporate page with words like safety, responsibility and trust. Ethics as the lived behavior of the model and the company behind it.

Anthropic's ethics often feels like paternalism. It may come from sincere concern. It may prevent real harm. But in daily work, it can also feel like the model believes it is your supervisor. That is not a small thing if you are building agents to run real workflows.

OpenAI's ethics feels more ambiguous: powerful, broad, increasingly infrastructure-like, but also shaped by platform politics, enterprise pressure and a very complex relationship with the public.

xAI's ethics is different. More freedom. More directness. More willingness to offend. More alignment with truth-seeking as a stated ideal. But also more exposure to Musk's temperament, X's chaos, and the concentration of power inside one man's industrial empire.

None of these options is pure.

That is the point.

When a model becomes part of your company and your life, ethics is not an abstract preference. It is operational.

Will the model let you work?

Will it refuse normal tasks because a policy layer panics?

Will it distort answers to fit institutional comfort?

Will it protect your sovereignty or slowly replace your judgment with its defaults?

Will it be available when ten agents need to run?

Will the provider punish power users for using the product exactly as the future demands?

The model is not just intelligence. It is a relationship with a worldview.

That is why I care about this beyond benchmarks.

why the power users matter

Anthropic may have thought it was protecting the serious market by prioritizing business customers.

But power users are not a distraction from the business market. They are the scouts.

They find the workflows before enterprises have names for them. They discover where agents break. They build weird integrations. They tolerate rough edges. They turn subscriptions into operating systems. They reveal what the model is actually good for.

Kicking them out may save compute in the short term.

It also sends them to competitors.

And once a power user rebuilds his stack around another model, he does not come back just because your benchmark improves by five percent. His tools, prompts, expectations, evals, habits and trust have moved.

This is what happened to me.

Claude was king. Then I had to route around Anthropic. GPT-5.5 became the default. Sol became better. Grok 4.5 became viable. The switching cost went down because Hermes itself became better. The harness improved. The agents improved. The model became one component in a stronger operating system.

That is the danger for model companies.

If the harness improves faster than the model gap widens, model loyalty weakens.

The intelligence becomes modular.

And when intelligence becomes modular, price, speed, access and philosophy become decisive.

google is the quiet exception

There is one caveat: Google.

At the beginning of the year I expected Google to step harder into the race. It has not done so in the loud way OpenAI, Anthropic and xAI have. But Google is not absent. It is playing a different game.

Google has the infrastructure. It has TPUs. It has distribution. It has research depth. It has data. It has a business that can subsidize intelligence in ways almost nobody else can.

If the game becomes factory speed, Google can compete.

Maybe it is quiet because it is behind. Maybe it is quiet because it does not need the same public theater. Maybe Gemini becomes the second player not because it wins Hacker News for a week, but because Google can industrialize intelligence across its own stack.

That is worth watching.

But the current momentum belongs to xAI.

Not because Grok is obviously the best mind today.

Because xAI looks like it is building the fastest learning machine.

the model is becoming a utility

This is the uncomfortable conclusion for the labs.

Frontier intelligence is becoming more like electricity.

We will still care about quality. We will still notice differences. Some models will have better taste. Some will code better. Some will write better. Some will reason more deeply. Some will be safer. Some will be freer.

But for daily work, acceptable intelligence is becoming abundant.

When that happens, users stop worshipping the model and start asking utility questions.

How fast is it?

How expensive is it?

How many agents can I run?

How often does it refuse?

How quickly does it improve?

Can I trust the company not to cut me off once I build around it?

Does this model make me more sovereign, or more dependent?

That last question is mine.

Because I do not use these models as toys. I use them to build software for AI4 Accountancy. I use them through Hermes to run a company and a life. I use them to think, write, decide, execute and coordinate.

So I do not only need a clever model.

I need a partner in a living system.

A model that is brilliant but scarce is useful.

A model that is nearly as brilliant, faster, cheaper and integrated into a compounding production machine may be more important.

That is the shift.

The frontier race is moving from model intelligence to intelligence production.

And if that is true, the winner will not be the company with the most beautiful demo.

The winner will be the company whose machine learns fastest.