When ChatGPT 5.5 Pro Does PhD-Level Math in an Hour, the Service Business Is Selling That Hour to Every Company That Can't Hire a PhD

Cambridge mathematician Timothy Gowers just published something that should rearrange your business thinking.

He tested ChatGPT 5.5 Pro on a set of open problems from additive number theory. Problems no one had solved. Problems serious academic mathematicians had not cracked.

Here is what happened.

In 17 minutes and 5 seconds, the model produced a construction that yielded a quadratic upper bound on a set diameter problem — which Gowers describes as "clearly best possible."

Then it improved a different bound three times in succession. The third improvement — a polynomial dependence for each fixed h — took 13 minutes, 33 seconds of thinking time. A collaborator who reviewed the output said the main idea was "original and clever" and the result "almost certainly correct" at both technical and conceptual levels.

Gowers's summary: "PhD-level research in an hour or so, with no serious mathematical input from me."

His conclusion about what this means: "It is no longer enough that somebody asks a problem: it needs to be hard enough for an LLM not to be able to solve it."

That sentence is not just about mathematics.

What just changed

The research layer — the part of intellectual work that involves surveying existing knowledge, identifying gaps, and generating novel approaches — has historically required expensive, credentialed, slow humans.

That was not a coincidence. Research is hard. Domain expertise takes years to build. The synthesis required to spot a gap and propose an approach is genuinely difficult.

But if ChatGPT 5.5 Pro can produce PhD-level output in under an hour on a cold start, the cost of that synthesis just collapsed.

Not to zero. The model still needs a human to validate the output, frame the question correctly, and turn the result into something the client can use.

But the hours of raw research work?

Those compress dramatically.

The business that exists right now

Most companies that need expert research are not buying it from academics. They are buying it from:

• consulting firms (slow, expensive, relationship-gated)
• analyst subscriptions (generic, not specific to their situation)
• freelance researchers (variable quality, hard to find, hard to evaluate)

None of those are AI-native.

The gap is an AI-native Research Sprint service.

Here is what it looks like in practice:

The client brings a specific question. Not "tell me about the AI landscape." Something like:

• "Which competitors are building AI-native features in our product category, and what are they actually shipping?"
• "What does the regulatory landscape look like for AI-generated medical records in the EU and the US right now?"
• "What are the strongest counterarguments to our pricing model, and what evidence exists for each?"

You run the frontier model — ChatGPT 5.5 Pro, Claude Opus, Perplexity — against the question with a structured prompt stack. You verify the outputs against primary sources. You package the result into a 3–5 page brief with citations, a one-page executive summary, and a set of follow-on questions the client probably has not thought to ask.

Turnaround: 24–48 hours.

Price: $500–$2,000 per sprint depending on complexity and domain.

The client pays for the output, not the hours. You capture the margin that used to fund junior researchers.

Why this is not just a prompt-wrapping business

The hard part of the Research Sprint is not the AI call.

It is knowing what question to actually ask. It is recognizing when the model is confidently wrong. It is having enough domain knowledge to validate the output against what is actually true. It is structuring the deliverable in a format the buyer can act on.

Those are human skills that become more valuable, not less, when the raw research capacity is cheap.

Gowers had to understand additive number theory to evaluate whether the model's answer was actually correct. His collaborator Rajagopal had to read the proof to call it "almost certainly correct."

Your Research Sprint business needs the same thing. A human who knows enough to check the work.

That is also your competitive moat. Anyone can run a frontier model. Not everyone can validate what it produces.

The best customer profiles

Research Sprints sell best to buyers who:

• already pay for research but find it slow or too generic
• face a specific deadline (a board meeting, a launch, a regulatory filing)
• are in a domain where being wrong is expensive (legal, medical, financial, regulatory)
• want a second opinion on analysis they have already done internally

Those buyers exist in every industry.

Good starting verticals:

Legal: Regulatory change monitoring, precedent searches, jurisdiction mapping. Law firms and in-house counsel already buy research. They understand what it costs.

Life sciences: Literature synthesis, competitive pipeline analysis, clinical trial landscape reviews. The research is well-defined and the clients have budget.

Enterprise strategy: Competitor intelligence, technology landscape assessments, market entry analysis. Buyers are internal strategy teams who need external perspective on a tight timeline.

Finance: Sector-specific due diligence research, regulatory filing analysis, earnings call synthesis. The outputs are structured and verifiable.

The retainer that follows

The natural upgrade from a Research Sprint is a standing research retainer.

One sprint a month. A coverage area the client defines. A recurring brief on the things they need to track but do not have time to monitor: competitor moves, regulatory changes, scientific developments, new entrants.

The model does the monitoring. You do the curation and validation. The client gets an expert-level briefing without hiring an analyst.

That retainer is worth $2,000–$5,000 per month for a serious buyer. It is completely defensible because it combines AI speed with human judgment in a format that would take the client weeks to replicate internally.

The bar just moved

Gowers says something important about what this moment means for researchers entering the field: "the bar has just been raised."

Problems that used to be publishable because they were open are no longer enough.

That is a problem for academics.

For service businesses, it is an opportunity.

The bar for research-adjacent services just dropped on the supply side while the demand from companies that need good analysis has not moved. The gap between "how much research costs to produce" and "how much buyers will pay for it" is wider than it has ever been.

The Research Sprint is how you step into that gap.

Sources:

https://gowers.wordpress.com/2026/05/08/a-recent-experience-with-chatgpt-5-5-pro/

https://news.ycombinator.com/item?id=43939120