Anthropic says two of its models, working with “minimal human involvement,” designed protein binders that were experimentally validated by external labs Adaptyv Bio and Twist Bioscience. Across fifteen targets, designs from Claude bound successfully 22 to 35 percent of the time, against what Anthropic describes as a 10 to 15 percent typical rate in the field — though that comparison figure is worth reading carefully, and this piece explains why below. Fourteen of the fifteen targets yielded at least one working binder, and on one target, Claude’s best design beat the winning entry from a human expert competition. The results come from an internal, non-peer-reviewed Anthropic report, not a peer-reviewed publication.

Why it matters

If the numbers hold up under outside scrutiny, this is a concrete data point that a general-purpose AI model — not a specialised structural-biology system — can do useful work in early-stage drug discovery, a step that normally takes trained scientists weeks per target. For pharmaceutical and biotech buyers evaluating AI-assisted discovery tools, it is a signal to start testing rather than a reason to change procurement plans: the study is small, internally run, and not yet peer-reviewed.

What Anthropic published

Anthropic’s own account, posted to its research blog on 18 August 2026 under the title “How Claude is accelerating protein design and analytical chemistry,” describes a project in which Claude autonomously called on two of its models — Mythos Preview and Opus 4.8 — to design de novo protein binders against a set of targets, then sent the designs out for physical testing. A third model, Opus 5, was used elsewhere in the same report for analytical chemistry tasks, not binder design. Anthropic’s post states plainly that “life science research tasks are currently blocked in our most capable model,” meaning the binder-design work was done with Mythos Preview and Opus 4.8, not with Anthropic’s current flagship.

Anthropic frames the work as a benchmarking exercise: could Claude models, with limited scaffolding and little human intervention, produce protein sequences that fold and bind in reality rather than only in simulation? Anthropic’s own materials describe this as a research report, not a peer-reviewed paper or a formal preprint.

Who actually tested the binders

The attribution claim here — that Claude “designed working protein binders” — rests on physical validation performed by two outside companies, and both have published their own accounts rather than only appearing in Anthropic’s telling.

Adaptyv Bio, which runs an automated wet-lab platform for protein testing, published its own case study on 19 August 2026. Adaptyv’s numbers: across 1,320 total designs, 354 bound their target, for an average hit rate of 26.8 percent, with 95 percent of designs expressing correctly. Sixteen targets were originally selected for testing, but one, GDF-8, produced no interpretable measurement at either outside lab and was excluded from the analysis set, leaving fifteen targets in the numbers both companies report — of those fifteen, one, MBP, produced zero binders, which is where the “14 of 15” figure comes from. Adaptyv also independently confirms the RBX1 example: Claude’s design improved on the human competition-winning entry, moving from a dissociation constant of 25.7 nanomolar to 3.9 nanomolar, a lower number meaning tighter binding.

Twist Bioscience, a DNA synthesis and biology-services company, published its own account describing itself as “an independent external evaluator” selected by Anthropic. Twist evaluated 1,260 protein binders across 15 targets in under three weeks. Twist did not publish detailed methodology for its own hit-counting, but its participation as a named, separately branded party is independent confirmation that a real outside lab ran physical tests on Claude’s output — though it confirms scale and role rather than the specific hit-rate percentages, which trace to Adaptyv’s numbers alone.

Both companies stand to benefit commercially from association with a high-profile Anthropic result, which is worth keeping in mind when weighing their statements.

What the metric actually measures

The headline “22 to 35 percent” is a hit rate — the share of designed protein sequences that were synthesised, expressed in a test system, and shown by surface plasmon resonance to bind physically to their intended target at a measurable strength. It is not a measure of whether any of these binders are safe, drug-like, or clinically useful; a binder that grips its target in a dish is an early-stage building block, several steps removed from anything resembling a drug candidate.

The comparison figure — that “10 to 15 percent is typical” in the field — is cited by Anthropic to Proteinbase, a protein-design-campaign database. But Proteinbase is not an independent third party: it is created and operated by Adaptyv Bio, the same company running the wet-lab validation for this study. Readers should treat the 10 to 15 percent figure as a comparison baseline sourced to an interested party, not as an independently audited industry benchmark.

What this means for buyers

Pharma and biotech teams evaluating AI-assisted discovery tools should treat this as a promising internal benchmark with two independent labs confirming that physical testing occurred, not as validated clinical-pipeline evidence. The work used Mythos Preview and Opus 4.8, not Anthropic’s most capable current model, and covered a narrow set of fifteen protein targets with binding-only measurements. Anyone building this into procurement or research decisions should ask Anthropic, Adaptyv, and Twist for the full target list, the raw data, and the exact scaffolding used, rather than relying on the headline percentages.

What would change our reading

A peer-reviewed publication or a preprint with full methodology and negative results would substantially strengthen the claim. So would an independent citation for the “10 to 15 percent industry typical” figure, and an independent replication by a lab with no commercial relationship to Anthropic, Adaptyv, or Twist.

Sources

  • Anthropic, “How Claude is accelerating protein design and analytical chemistry,” 18 August 2026 — anthropic.com
  • Adaptyv Bio, “Case study: Benchmarking Claude’s protein designs in the wet lab,” 19 August 2026 — adaptyvbio.com
  • Twist Bioscience, Anthropic AI protein design resource page — twistbioscience.com