Anthropic Estimates AI's Potential in Science: Up to 2x Now and Up to 15x in the Future

A researcher usually doesn't have a single job but a chain of dozens of different tasks: finding literature, understanding someone else's code, writing your own, preparing data, checking calculations, describing the result, and only then returning to the experiment itself. AI is good at speeding up the digital segments of this chain, but it cannot automatically remove all the other constraints.

Anthropic analyzed more than 207,000 sessions of Claude and Claude Code users related to scientific work. The company's conclusion sounds stronger than the usual "scientists have started using chatbots more often": with AI's current capabilities, the volume of results in certain fields could theoretically almost double. And if it ever becomes possible to automate the entire suitable cycle, the gain could reach 15x.

It is important not to confuse the two figures. The nearly twofold speedup is an estimate of the effect of tools available today at the stages where a model can already help. The fifteenfold growth is a forward-looking scenario of full automation, not a result a research group gets after installing Claude Code.

Where the speedup comes from

Imagine a small team testing a new hypothesis. Previously, a researcher spent two days searching for suitable libraries, another day assembling a data-processing script, and several hours fixing errors. AI can shorten this part of the work: suggest a starting point, explain unfamiliar code, write a draft script, and help with debugging.

But if the sample then has to be fabricated, placed in an apparatus, and observed for a week, the overall project will not become ten times faster. Work at the laptop got faster, yet the physical stage stayed the same. This is a familiar engineering picture: optimizing one segment is not the same as speeding up the whole system.

That is why the practical value of Claude and similar tools depends not only on the quality of answers. It is determined by what share of the process can be handed to the model, how much time it takes to verify its work, and where the real bottleneck lies. For computational research the effect may be more noticeable than for laboratory disciplines, where results depend on equipment, materials, and human presence.

Why 15x should not be read as a promise

The study is built on an analysis of user sessions, not on multi-year observation of identical teams with and without AI. Such data shows well which tasks people hand to the model, but by itself does not prove that overall scientific productivity has grown by the stated factor.

There are other limitations. The amount of code created, data processed, or texts prepared is not yet equal to the number of quality discoveries. A model's error can speed up the production of an incorrect result. The more important the conclusion, the more expensive independent verification, reproducibility, and expert review become. Finally, the analysis covers users of Anthropic's products, so their behavior cannot automatically be extended to all scientists.

The 15x estimate is more sensibly treated as the upper bound of a scenario: what happens if AI learns to reliably perform almost all digital tasks and the processes around it are rebuilt. Between such a model and today's laboratory lie integration with tools, quality control, data access, security, and accountability for conclusions.

What developers and product teams should do

The main practical takeaway is not to start with the promise of "we will automate research end to end." First, break the process into stages and measure each one: source search, code writing, data analysis, verification, experiment, and publication preparation. Then pick a segment with a lot of repeatable digital work where the result can be checked quickly.

For a product, it is more useful to count not the number of model queries but the time to a verified result, the share of human corrections, and the rate of detected errors. If AI writes code twice as fast but the researcher spends the same amount of time checking it, there may be no real gain.

The Anthropic study does not prove that science will speed up 15x tomorrow. It shows something more useful: AI's potential is determined by the architecture of the workflow. Teams that find a measurable bottleneck and build verifiable automation into it will benefit right now. Everyone else will get an impressive number with no change in results.