Fable 5.1's Molecular Design Capability Isn't That Magical

Anthropic just released Fable 5.1, and it showcased quite a few biomedical and molecular-design capabilities.

When you see this kind of promotional material, the first question worth asking is: does this capability actually come from the LLM itself, or does it come from an agent calling external tools?

This distinction matters. We shouldn't expect that feeding in an amino acid sequence gets you a reliable 3D structure straight out of the LLM. The actual work is usually still being done by dedicated models like ESMFold, AlphaFold, or Boltz — the LLM is mostly acting as an agent, responsible for calling, orchestrating, and deciding the next step.

So the evaluation standard is simple: if you swapped in a different strong LLM, could this feature still be done?

Unsurprisingly, Anthropic is once again drawing peptides. Per their official announcement, there are roughly four areas of progress:

1. Agentic protein design

This direction is actually fairly mature by now.

An LLM calls structure-prediction and protein-design models, generates a binder, then sends it off for experimental validation. As long as you're willing to spend the money, a full closed loop is entirely doable.

One peptide chain plus some basic experiments — that's a few thousand to ten thousand RMB for synthesis plus validation. And the ones loudly advertising "automated generation and synthesis" are priced several times higher than any random peptide synthesis company in the Zhangjiang tech park. These days even a high schooler using ChemOrchestra can design a batch of candidates in one pass, and plenty of downstream CROs are happy to take the work. So I don't think the real "closed loop" is landing a Hit. It's:

design → experimental validation → optimization → and finally, license-out — actually making money from it.

Only once you've done that does a drug-design agent actually count as having entered the industry. And I wouldn't recommend people go after peptides/proteins specifically, because they're genuinely hard to sell. Peptides are inherently "sticky" — easy to land a hit on. We're currently planning a skincare-peptide design hackathon, which is a much better fit for actually bringing something to market.

2. Automatically optimizing protein models

Fable 5.1's promotional material also claims it can read biological model code itself, then optimize GPU kernels, caching, and execution paths. I think other coding agents can already do a fair amount of this today.

I've personally used leftover Claude Code tokens with Sonnet to optimize some large protein-prediction models. For a model like ESMFold2, doing fused kernels and reducing memory read/write in various places, getting a localized 10x speedup isn't unusual. What NVIDIA does with Boltz-2 NIM is essentially the same thing — pushing inference speed further via TensorRT and kernel optimization. Someone even migrated CUDA to the AMD ecosystem over a single weekend.

3. Hypothesis generation

I'm still fairly reserved on this one for now. A lot of hypotheses in biomedicine come from a person reading a large number of papers and guessing from scattered, fragmentary results.

And this field has an interesting quirk: a wrong hypothesis can sometimes stumble into a good result anyway.

You think the mechanism is A, and it turns out to actually be B — but the drug really does work. So a model proposing a mechanism that "sounds reasonable" — I think the value there is still fairly limited. What's needed at this stage is a human's wildly imaginative capacity to be wrong, not model hallucination.

4. LLM-controlled lab work

As long as liquid handling, mass spec, automated pipettes, and synthesis platforms all have open, stable APIs, then whatever Fable 5.1 can control, any other LLM can control too.

So this is really more of a lab-infrastructure question. In the end, a lot of nonstandard experiments still need to be done by hand — organic synthesis, for instance. That said, I've recently been connecting Claude to a 3D printer via MCP, and Claude has been quite careful handling support-material decisions.

What I'm actually most looking forward to from AI language models, though, is their math ability.

I've previously used Fable 5 together with Maple to work through new functionals, or things I'd had ideas about but could never fully derive because the math was too complex — Opus and ChatGPT weren't quite up to it either, and I kept getting stuck partway through. If models genuinely keep pushing this kind of mathematical reasoning forward, I think that's actually more interesting than drawing a few more peptides nobody's going to buy. Fable 5.1 spending its effort drawing peptides that won't sell feels like a waste of real talent.

Follow us — we'll be putting out tutorials on building your own research agent, wired into your own lab's or company's proprietary knowledge, which will beat Fable 5.1 in that context, won't refuse to answer just because the topic is biology, and — most importantly — keeps your data secure.

If you want a walkthrough of the platform or have any specific question/needs, you can reply to this email or schedule a meeting with me: https://calendly.com/yizhouma/chemorchestra-user-session