A startup founded by DeepMind veterans just beat Anthropic and OpenAI at one of the hardest tasks in science: replicating research. Inherent says its AI teammate scored higher than both companies’ flagship models on scientific research replication benchmarks. If that holds up under scrutiny, it changes what “best AI” actually means.
Why This Matters Right Now
Scientific research replication is in crisis. According to a landmark survey published in Nature, more than 70% of researchers have tried and failed to reproduce another scientist’s results. That number has barely moved in a decade. The scientific community calls this the “replication crisis,” and it is costing real money.
According to a study by Freedman, Cockburn, and Simcoe published in PLOS ONE, irreproducible preclinical research costs the US biomedical industry an estimated $28 billion every year. That is not a rounding error. That is wasted drug trials, failed grants, and careers built on results nobody can verify.
Into this mess walks Inherent. The company was founded by alumni from DeepMind, the AI research lab owned by Google. Their product is an AI “teammate” built specifically to help scientists verify and replicate existing research. According to Inherent, their system scored higher than models from both Anthropic and OpenAI on standardized research replication benchmarks. That claim landed in 2026, when the AI market is already crowded and every company is fighting for attention with big announcements.
But this one is different. Here is why.
The Contrarian Take Nobody Is Saying
Everyone is chasing the same prize: the most capable general AI. OpenAI wants to be everything. Anthropic wants to be the safe version of everything. Google wants to be the integrated version of everything. They are all building wider and wider. Inherent went narrow.
That is the move most people miss.
Speed without accuracy is how you waste money at scale. The pharma industry knows this. A drug company can run a study faster than ever with AI. But if the foundation study they are building on cannot be replicated, they are spending millions on a broken premise. The speed makes the damage worse, not better.
Inherent trained their system on the specific patterns that make scientific work reproducible. Methods sections. Data formats. Statistical approaches. Variable controls. That is a completely different training objective than “answer questions well” or “summarize documents.” It is specialized. And according to Inherent’s published benchmarks, that specialization produced a model that outperformed Claude and GPT on the tasks that matter most to scientists.
I have been watching this pattern for a while now. The biggest names in AI are optimized for the broadest possible use cases. That means they are mediocre at the most specific problems. A DeepMind trained team that spent years on hard science knows exactly where that gap lives.
Here is the rich versus poor mindset applied directly to this moment. The average operator sees AI as one tool they plug in everywhere. The sharp operator asks: which specific pain point is expensive enough, specific enough, and neglected enough that a narrow solution wins? Inherent found all three. According to data from Grand View Research, the global scientific research services market was valued at over $1.7 trillion in 2024. You do not need a big slice of that market to build a serious business. You need the right slice.
The companies and universities that prove their research is reproducible will win grants faster, close pharma partnerships faster, and clear regulatory hurdles faster. That makes research verification a competitive weapon, not just a quality check. Whoever owns that tool owns serious pricing power.
If you create content or explainer videos about AI developments like this one, InVideo AI lets you turn articles into short videos without a production team. For a story this technical, video often reaches the audiences that never read past the headline.
What This Means for You
If you work in science, biotech, pharma, or academic research, start paying attention to Inherent. They are not widely deployed yet. But the category they are building in is about to become a requirement, not a nice to have.
Here is what I would do right now. If I ran a research team, I would map out which AI tools are being purpose built for scientific verification versus which general tools are being adapted after the fact. The accuracy gap between those two categories is real and it will get wider as specialized models keep improving.
If I ran a biotech startup or a pharma department, I would be asking my vendors one question: can your AI show its work in a way that a peer reviewer can verify? General AI tools mostly cannot. A purpose built system like Inherent is built around that exact requirement.
For founders watching this play out, the lesson is clear. Narrow beats broad when the problem is specific enough and expensive enough. Inherent found a $28 billion annual problem and built a focused tool for it. That is a real business with real pricing power.
For operators building content or media businesses around AI coverage, understanding niche breakthroughs before they trend is where the edge is. AppSumo carries lifetime deals on research tools, writing software, and publishing platforms that can help you move faster on stories like this before the mainstream press catches up.
The Bottom Line
Inherent just proved that being the best at one hard thing beats being decent at everything. A team of DeepMind veterans picked a $28 billion problem, ignored the noise, and beat the biggest names in AI at a benchmark that actually matters to an industry. The next wave of AI value is not coming from the companies with the biggest valuations. It is coming from the teams nobody has heard of yet. Inherent is one of them. Watch this space closely.
Frequently Asked Questions
What is Inherent AI?
Inherent is an AI company founded by alumni from DeepMind, Google’s AI research lab. Their main product is an AI “teammate” built specifically to help scientists replicate and verify research findings. The company focuses on one of science’s most expensive problems: the inability to reproduce published results.
How did Inherent beat Anthropic and OpenAI on research tasks?
According to Inherent, their AI scored higher than models from both companies on scientific research replication benchmarks. The advantage likely comes from specialized training on the specific patterns that make research reproducible, rather than broad general training. Narrow specialization tends to win on domain specific tasks.
What is the scientific research replication crisis?
According to a Nature survey, more than 70% of researchers have failed to reproduce another scientist’s results. According to research published in PLOS ONE, this costs the US biomedical industry an estimated $28 billion per year in wasted trials and bad decisions built on unverifiable findings.
Should I be using AI for scientific research verification right now?
If you work in pharma, biotech, or academic research, general AI tools are not built for this. Purpose built tools like Inherent are designed around reproducibility and verification. As this category matures, using the right specialized tool will matter for regulatory approvals, grants, and partnerships.
Does this mean narrow AI will beat general AI going forward?
In specific high stakes domains, yes. General models are optimized to be decent at everything, which means they are often mediocre at the hardest specialized problems. When the problem is specific enough, expensive enough, and neglected enough, a narrow purpose built model tends to win. Research replication is a clear example.
