OpenAI's 722 Math Results and Musk Picking Claude

OpenAI dropped a pile of new mathematics results in one day. They came from a model you cannot see or use.

Elon Musk, meanwhile, says his products will run on Anthropic's Claude Opus 5.5. Whenever that model gives the best outcome, it gets the job.

One company is advertising a brain it keeps locked up. The other is renting a competitor's brain and posting about it. Both are more interesting than they sound.

OpenAI's 722 math results and the model nobody can see

On October 6, OpenAI published a public collection of mathematical results, organized into 372 result families. The number people are repeating is 722 results. Either way, it is a lot of mathematics to produce in one go.

They pointed an internal model at a big batch of open problems. The list includes:

• Matrix multiplication
• Algorithmic optimization
• Classic combinatorics results
• Aperiodic 3D tiling

Some of these have been sitting there untouched for 50 to 60 years.

The model that did it has not been named. You cannot use it. You cannot inspect it. You get the papers and a tweet.

Now the part that matters. These are leads, not verdicts. Every result still needs a human expert to check it. Then someone else has to replicate it.

Hundreds of promising answers just landed on mathematicians' desks. They now have months of unglamorous work ahead. The machine did the fun part and outsourced the homework.

I still think this is a big deal, and the reason is boring. Mathematics has a clean right-or-wrong signal. A proof holds or it falls over.

That makes it a perfect training ground. You can reward the model automatically, with no human arguing about taste. Reasoning models are roughly two years old. Two years ago they were fumbling arithmetic.

Everyone is now predicting the same thing will happen to physics, chemistry and biology. Maybe. That is a prediction, not news. And biology does not come with a neat tick-or-cross answer sheet.

My own guess: fields with cheap, automatic checking move first. The rest move at the speed of lab work and regulators.

Here is what this is worth to you if you run a company with 12 employees and a messy order system. Nothing, today. But the pattern underneath is directly useful. AI gets good fastest at tasks where you can check the output automatically.

So when you pick a process to automate, pick the ones with a correct answer. Does the invoice total match the bank line. Does the extracted VAT number have the right format. Does the generated code pass the test.

Those jobs are safe to hand over. A wrong answer announces itself.

The jobs where nobody can tell whether the output is wrong are the ones that quietly cost you money. That is true for a frontier research model and for the little workflow you set up last Thursday.

Why Elon Musk is routing Grok to Claude Opus 5.5

Musk posted a note saying that going forward they will use the best back-end model for any given task. He named Claude Opus 5.5, Midjourney, Suno and other leading APIs. Whatever is most likely to give the best outcome. The post also mentioned SpaceX engineers doing the same thing internally.

Look at that list for a second. Midjourney makes images. Suno makes music. Claude Opus 5.5 from Anthropic is a frontier reasoning model.

Those are not equal options. Putting them in one sentence is a choice. It is like saying your restaurant buys from a fishmonger, a baker, and another restaurant that cooks everything for you.

The actual headline is that a man who builds his own models is routing serious reasoning work to a competitor's. And he said it out loud, in public, with no press release softening around it.

I respect it. It is the correct engineering decision, and it is the one I would make. Brand loyalty to a model provider is a tax you pay for no service.

The sensible setup is a router. Each task goes to whichever model does it best this month. That answer changes often.

There is also a cold business logic here. Musk already owns the distribution. X has the users. More importantly, it has a lot of users with real power.

If Grok becomes the thing people type into, the model behind it is just a supplier. Suppliers get swapped. That is the whole point of being the layer on top.

For your company, the lesson is smaller but more practical. Build so you can change models without rebuilding everything.

In practice that means keeping your prompts, tools and business rules separate from whichever API you called this week. If swapping provider means a weekend of rewriting, you did not buy an AI system. You bought a long-term relationship you did not agree to.

We think about this a lot when we build AI agents for clients. The best model for a job genuinely changes every few months.

And judge models on your own work, not on charts. Benchmark scores are produced by the people selling the model. Your invoice-matching accuracy is produced by your actual invoices. One of those two numbers pays your rent.

The quiet contradiction between the two

Put the stories next to each other and something funny falls out. OpenAI published hundreds of results and kept the model private. Great way to look powerful without letting anyone test the claim. Musk published nothing about capability and simply told everyone which model he pays for.

One of those is marketing. The other is a purchase order. I find purchase orders more honest.

If you want a rule of thumb from today: trust what companies buy more than what they announce. And if a lab tells you its secret model solved everything, ask who checked it.

My prediction for the next year is that nearly every serious AI product becomes a router with a logo on it. At which point the question stops being which model you use, and becomes whether anyone noticed you switched.