Anthropic says Claude speeds up 30+ biomolecular models by average 4x

What you will learn in this article

・In an article on its official website titled "How Claude is uplifting biomolecular modeling" (published September 17, 2026), Anthropic reported the results of an experiment using Claude to optimize over 30 open-source AI models in the life sciences field. The original text states, "Claude, working within Claude Science, optimized more than 30 of these models in just under four weeks, speeding them up roughly 4x on average."

・This optimization was supervised by two members of Anthropic's technical staff who have expertise in biomolecular modeling but no prior experience in inference optimization or kernel engineering. The original text states, "Claude, supervised by two members of Anthropic's technical staff who are experienced in biomolecular modeling but who had no prior experience in inference optimization or kernel engineering..."

・The "average 4x" figure is presented in the introduction of the original text as an average across all 30+ models spanning structure prediction, protein design, genomics, and protein language modeling, under conditions where precision is slightly sacrificed. Under conditions where outputs are identical, the speedup is "nearly 2x." This article treats this as a single experimental result conducted by Anthropic itself, not as a general guarantee for AI-driven code optimization.

"In this post, we share how Claude made the open-source models that scientists use to predict and design biomolecules faster and more memory-efficient."—Anthropic reported this on its official website.

Summary of primary information—What is the experiment where Claude optimized over 30 models?

The primary source for this article is the Anthropic official website article, "How Claude is uplifting biomolecular modeling" (published September 17, 2026).

The original text continues to describe the overview of this initiative: "Claude, working within Claude Science, optimized more than 30 of these models in just under four weeks, speeding them up roughly 4x on average."

Translated into Japanese, this means, "Working within the Claude Science environment, Claude optimized over 30 of these models in just under four weeks, speeding them up by an average of roughly 4x."

In other words, this experiment is a case study where Claude was used to perform code optimization tasks within "Claude Science," an AI workbench for researchers announced by Anthropic in June 2026 (also covered in previous AI TREND LAB articles).

Regarding the supervision structure, the original text explains more specifically: "Claude, supervised by two members of Anthropic's technical staff who are experienced in biomolecular modeling but who had no prior experience in inference optimization or kernel engineering, carried out the acceleration of more than 30 open-source models across biomolecular structure prediction, protein design, protein language modeling, and genomics in just under four weeks."

Translated into Japanese, this means, "Under the supervision of two Anthropic technical staff members who have experience in biomolecular modeling but no experience in inference optimization or kernel engineering, Claude executed the acceleration of over 30 open-source models spanning biomolecular structure prediction, protein design, protein language modeling, and genomics in just under four weeks."

In other words, it is explicitly stated that the two people who supervised this experiment have professional knowledge in the target field (biomolecular modeling) but lacked prior professional experience in the actual work Claude performed (inference acceleration and GPU kernel optimization).

Core facts—The precise meaning of "average 4x" and the target models

Regarding the overall picture of the acceleration, the introduction of the original text explains: "Claude was able to optimize more than 30 deep learning models trained for a variety of biological tasks, such as structure prediction and protein design, as well as for genomics and protein language models."

Translated into Japanese, this means, "Claude was able to optimize over 30 deep learning models trained for various biological tasks, such as structure prediction and protein design, as well as for genomics and protein language models."

The original text continues: "On average, Claude was able to speed up such tasks roughly 4x while sacrificing a minimal amount of precision, and nearly 2x with identical outputs."

Translated into Japanese, this means, "On average, Claude was able to speed up these tasks by roughly 4x under conditions that sacrifice a minimal amount of precision, and by nearly 2x under conditions where the output is completely identical."

In other words, the introduction of the original text clearly states that the "average 4x" (with minimal precision sacrifice) and "nearly 2x" (with identical outputs) are figures covering all 30+ models across the four categories of structure prediction, protein design, genomics, and protein language modeling.

Separately, in the "Accelerating structure prediction models" section of the original text, there is a specific description limited only to structure prediction models: "The combination of these improvements accelerated the structure prediction models by 4x, on average, and for each model, we confirmed that Claude's accelerated versions did not impact performance on the downstream task (such as structure prediction)."

Translated into Japanese, this means, "By combining these improvements, the structure prediction models were accelerated by an average of 4x. And for each model, we confirmed that Claude's accelerated versions did not impact performance on downstream tasks (such as structure prediction)."

In other words, the original text contains two descriptions: (1) the "average 4x / nearly 2x" figures covering all 30+ models across four categories, and (2) the same "4x" figure that happens to apply only to the structure prediction models.

The "average 4x" used in the title and headings of this article refers to the overall average in (1), and the extent to which each category contributes to this overall average (the breakdown by category) cannot be determined from the description of the primary information checked this time.

Regarding memory efficiency, the original text provides specific figures: "Claude created a low-memory 'Big' mode that enables the accurate modeling of systems larger than 10,000 tokens and successful inference on systems larger than 70,000 tokens using just one NVIDIA GPU node—a previously out-of-reach task."

In Japanese, this means: "Claude created a low-memory 'Big' mode, enabling accurate modeling of systems larger than 10,000 tokens and inference on systems larger than 70,000 tokens using a single NVIDIA GPU node."

The original text positions this as "a previously out-of-reach task."

Regarding specific examples of the targeted models, the original text states: "Modern structure prediction models, such as AlphaFold3, OpenFold3, and Boltz-2, spend much of their computational runtime and memory on two operations: triangle attention and triangle multiplication, which act on triplets of tokens."

In Japanese, this means: "Modern structure prediction models such as AlphaFold3, OpenFold3, and Boltz-2 spend much of their computational runtime and memory on two operations: triangle attention and triangle multiplication, which act on triplets of tokens."

Code release—36 optimization kits available on GitHub

The original text also explicitly mentions the release of the optimization code: "We are open-sourcing the optimized code for all of these models today (here) so that the broader community can make use of them."

In Japanese, this means: "Today, we are open-sourcing the optimized code for all of these models so that the broader community can make use of them."

This code can be found in the GitHub repository 'anthropics/uplifting-biomolecular-modeling'.

Upon checking the repository, I found that 36 inference optimization kits (including af2ig, af3_jax, af3_torch, atlasfold, boltz2, chai1, colabfold, esmfold2, openfold3, rfdiffusion1, rosettafold3, and others) have actually been released, confirming a more precise breakdown that is consistent with the 'over 30' figure mentioned in the blog post.

Each kit provides multiple modes that appear to correspond to the 'conditions that slightly sacrifice accuracy' and 'conditions where the output is identical' mentioned in the original text (off = original version without optimization, exact = faster with identical output to the original, fast = faster while allowing for slight numerical differences, big = supports large-scale inputs with minimal GPU memory).

The repository explicitly states 'not maintained and not accepting contributions,' indicating that it is intended as a reference implementation accompanying the blog post.

Regarding licenses, I confirmed that the original optimization code released by Anthropic is under the Apache License 2.0 for each kit.

For each of the open-source models (upstream projects) targeted, the GitHub README includes a table with an 'upstream licence' column, specifying the individual license for each kit.

For example, atlasfold and boltz2 are MIT, af3_jax is Apache-2.0, and rfdiffusion1, rosettafold3, and progen2 are BSD-3-Clause.

Note that af2ig does not have a single license but is listed as 'MIT; AlphaFold 2 code Apache-2.0,' indicating multiple licenses where the kit's own code is MIT and the incorporated AlphaFold 2-derived code is Apache-2.0.

It was confirmed that licenses differ depending on the upstream model and, in some kits, the origin of the incorporated code.

While I attempted to check the technical report (PDF) linked from the original text during the production of this article, the file size exceeded the limit of the retrieval tool, so I could not directly verify its contents.

This article is based on information available from the official blog post and the GitHub repository.

A perspective often overlooked—there are many conditions that cannot be generalized

There are a few points to keep in mind.

First, this experiment was conducted and reported by Anthropic itself, and has not been verified by an independent third-party organization.

Second, the figures of '4x on average' and 'approximately 2x' are explicitly stated in the introduction of the original text as the average across more than 30 models spanning structural prediction, protein design, genomics, and protein language modeling.

On the other hand, the original text also contains a separate individual description of '4x on average' targeting only structural prediction models, and it is impossible to determine from the primary information confirmed this time how much each category contributes to this overall average (the breakdown by category).

Additionally, it is explicitly stated that while the two supervisors had expertise in biomolecular modeling, they had no experience in inference optimization or kernel engineering.

Conversely, while this is a case study showing that Claude itself worked to supplement optimization knowledge in this field, it should be noted that the supervision of humans with expertise in the target field is a prerequisite.

Furthermore, I would like to ensure that the difference in conditions is not overlooked: 'approximately 4x' is a figure under conditions that slightly sacrifice accuracy, and it remains at 'approximately 2x' when seeking output identical to the original model.

The '4x average' in the title of this article must also be understood based on the fact that it is a conditional figure.

Finally, what was optimized this time was a specific group of open-source models in the life sciences field, and it cannot be determined from this primary information alone whether these results can be generalized to AI-based code optimization in general (all types of software and all acceleration tasks).

Implications for developers, AI utilization managers, and individual developers

From here on, this is the perspective of AI TREND LAB based on the primary information, and I would like to state in advance that it includes speculation rather than definitive conclusions.

For developers, the fact that the use of generative AI is expanding beyond new code generation to performance optimization of existing open-source software and AI models themselves seems to be a useful reference.

However, this case was conducted under a system where two humans with expertise in the target field supervised it, and it cannot be confirmed from this primary information whether similar results can be reproduced by supervisors without expertise.

For AI utilization managers, when considering performance improvements for software and AI models used in-house, it seems worth considering the direction of 'having AI optimize existing code' as an option other than 'replacing with new tools'.

In that case as well, it seems that establishing a system where humans with expertise in the target field verify the results, as in this experiment, will be a prerequisite.

For individual developers, the optimization code and methods published on GitHub this time (such as the design of multiple modes like exact, fast, and big) can serve as a reference case when entrusting code optimization to AI in their own projects.

However, it should be noted that the repository itself is a reference implementation that is not being maintained.

Summary: Things you can try starting tomorrow

What this experiment shows is the fact that Anthropic, within an environment called Claude Science, had Claude optimize over 30 open-source AI models in the life sciences field, achieving an average 4x speedup, albeit under certain conditions.

It is necessary to be aware that this figure is merely an experimental result under specific conditions (settings that slightly sacrifice accuracy, supervision by humans with expertise) and is not a guarantee for AI-based code optimization in general.

As something you can try starting tomorrow, for the software or AI models you or your company maintain, you might want to consider the option of "AI-driven optimization of existing code" in addition to "replacing them with new tools."

In doing so, it would be wise to also plan for a system where humans knowledgeable in the target field verify the results, just as in this case study.

For details, please be sure to check the primary sources directly.

Source:

Anthropic, "How Claude is uplifting biomolecular modeling" (Anthropic official website, September 17, 2026)

https://www.anthropic.com/research/claude-uplifts-biomolecular-modeling

Anthropic, "uplifting-biomolecular-modeling" (GitHub repository)

https://github.com/anthropics/uplifting-biomolecular-modeling

Conclusion:

At AI TREND LAB, we carefully verify research, reports, and official announcements released overseas using primary sources. We translate and organize this information from the perspective of how Japanese business professionals can apply it to their practical work or side projects, while avoiding exaggeration or definitive claims.

Moving forward, AI TREND LAB will continue to provide accurate and easy-to-understand updates on the latest AI trends based on primary overseas information.