Claude Sped Up 30 Biology Models 4x — and Anthropic Opened Verified Access

Anthropic published a research post on September 17, 2026 showing that Claude, working inside its internal Claude Science effort, accelerated more than 30 open-source biomolecular models by roughly 4x on average — and opened public beta applications for a new Life Sciences Verification Program that gives vetted labs and biotechs access to its models, including Mythos, for biology-related work. Alongside both announcements, Anthropic is co-sponsoring a protein design competition with Adaptyv Bio backed by up to $1 million in Claude credits. Here's what Claude actually built, what it means for the cost of doing computational biology, and how the new access program works.

TL;DR

The actual engineering: FlashPairformer

Modern structure-prediction models — AlphaFold3, OpenFold3, Boltz-2 — spend most of their runtime and memory on two operations, triangle attention and triangle multiplication, which act on triplets of tokens to model biomolecular geometry. Both are cubic in cost: tripling the size of a system uses 27x more time and memory. That cost curve is the reason large biomolecular systems have historically required either heavy compute budgets or multi-GPU-node inference that's out of reach for most academic labs.

Anthropic worked with Claude to build FlashPairformer, custom kernels targeting exactly those two operations. Anthropic reports it outperforms the field standard by 2.7-2.9x on triangle attention and 1.7-3.2x on triangle multiplication, depending on model configuration — a "field standard" that includes NVIDIA's own BioNeMo Inference Runtime, which explainx.ai covered when it entered public beta on September 10, 2026. Beyond the transferable kernel work, Claude was also pointed at more than 30 individual models to find model-specific optimizations — caching redundant computation, simplifying dead branches into constant outputs — and Anthropic reports the combination sped up structure-prediction models by 4x on average.

The organizational detail is arguably the more interesting claim: Anthropic says this work, which "normally takes an experienced team of engineers weeks to produce... and the work often does not transfer between models," was carried out by Claude in under four weeks, supervised by two Anthropic staff who had biomolecular modeling experience but no prior background in inference optimization or kernel engineering. That's a claim about what a domain expert paired with an agentic model can now ship without also being a systems specialist — read alongside Anthropic's earlier work on Claude proving Fermat's Last Theorem in Lean as another data point in the same pattern: Claude doing sustained, autonomous technical work over multi-week horizons with a domain expert supervising rather than driving.

"Big" mode: modeling systems that didn't fit on one GPU before

Alongside speed, Claude built a low-memory "Big" mode that enables accurate structure prediction for biomolecular systems larger than 10,000 tokens — amino acids, nucleotides, and small-molecule or ion atoms — on a single NVIDIA GPU node, and successful (if not always accurate) inference beyond 70,000 tokens. Anthropic reports molecular machines folded successfully using Big mode include a human mitochondrial complex I, the TRiC chaperone complex, a proteasome, and a bacterial ribosome, each closely matching experimentally determined structures — among the largest structures Anthropic says have ever been folded accurately this way, well beyond AlphaFold3's previously reported 7,663-token 40S ribosome prediction.

Pushed further, using a single 8-GPU B300 node, Claude generated predictions for entire viral capsids and protein compartments ranging from 31,000 to over 70,000 tokens — nearly two orders of magnitude beyond these models' training context. Anthropic is explicit that those predictions are not correct at that scale ("predicted structures collapse, suggesting a lack of generalization"), but frames the achievement as lowering the access barrier: what previously required multiple GPU nodes now runs on one, even if accuracy at the extreme end still needs work.

The cost collapse: $10,000 per target to $150

The most concrete number in Anthropic's post is a cost comparison against its own earlier work. In a previous de novo protein binder design campaign covered on explainx.ai, Anthropic gave Claude an approximately 16,000-word prompt, sub-agent access, and up to $10,000 per target (roughly 2,500 NVIDIA H100 GPU hours) across a 24-hour window — and hit a 22-35% success rate against a field norm of 10-15%.

With the new accelerated models, Anthropic gave a single Claude model access to one NVIDIA H200, 24 hours of wall time, a roughly 1,100-word prompt, and no sub-agents or human steering. Run against 16 targets and scored by ipSAE (an in silico score shown to predict wet-lab binding), the median and highest-scoring designs from three Claude models (Mythos 5.1, Mythos 5, Opus 5) matched the earlier campaign's in silico performance — using about two orders of magnitude fewer GPU hours. Combined GPU and token spend: roughly $150 per target. That's not a marginal efficiency gain; it's the kind of cost collapse that changes who can afford to run this kind of design campaign at all, from well-funded pharma R&D budgets to individual academic labs.

The Life Sciences Verification Program: two grant types, three threat models

Separately, Anthropic opened public beta applications for the Life Sciences Verification Program (LSVP) on September 17, 2026, giving vetted researchers access to Mythos, Opus, and Sonnet models for work currently blocked under Anthropic's generally available Fable models — drug discovery, research biology, clinical development, manufacturing. Anthropic says it has already onboarded dozens of organizations through an early-access period before opening broader applications, and expects to enroll hundreds more within the first week.

Verification runs against research credentials, security standards, and ethical research oversight, and approved teams get one of two grant types:

• Standard Use — covers most life-science work (basic science, R&D, supply chain, manufacturing, clinical development, QA, regulatory affairs, and more), extends to whole teams, renews annually, and applies today to Mythos 5.1, Opus 5, and Sonnet 5.
• High-risk Use — an add-on for dual-use work that Standard Use blocks, such as characterizing how a specific viral vector family is recognized by human immune pathways. It removes safeguards for that specific project only (not the whole team), renews every six months, and for Mythos specifically remains limited to a small set of vetted entities working with the US government — Opus 5 and Sonnet 5 high-risk grants are available today.

Anthropic frames the design around three threat models it's explicitly trying to defend against, rather than a generic "prevent misuse" goal: access compromise (malware or account takeover redirecting legitimate access to a bad actor), insider threats (a rogue or coerced employee diverting access), and agent misuse (agents, especially operating in swarms over long-horizon tasks, taking unintended dangerous actions on their own). The enforcement mechanism shifts from real-time request blocking toward offline monitoring — LSVP traffic is retained for 30 days specifically to let Anthropic detect misuse patterns spread across many sessions, data that's compartmentalized from model training and from Anthropic's own life-sciences research teams. Availability today is limited to Enterprise/Team console, Claude Science, and Claude Code with API auth — individual Pro/Max plan access and BAA-enabled (HIPAA) orgs aren't supported yet. This graduated, monitored-access structure is the concrete implementation of the kind of embedded-evaluator thinking explainx.ai covered in Anthropic's broader "Pace the Frontier" safety commitments.

The Adaptyv Bio competition

To put the accelerated models to a public test, Anthropic and Adaptyv Bio are co-sponsoring a protein design competition covering five challenge problems at the frontier of current capability — species cross-reactivity, pH-sensitivity, peptide-MHC specificity, and difficult GPCR targets among them. Anthropic is committing up to $1 million in Claude credits plus additional funding for wet-lab validation; Modal is providing up to $250,000 in compute credits; Twist Bioscience is providing DNA synthesis. Anthropic and Adaptyv say they'll experimentally validate more than 5,000 community-submitted designs and share all results openly.

Honest limitations

• These are Anthropic's own self-reported benchmarks, run and scored by Anthropic's own methodology, with no independent third-party audit disclosed alongside the announcement.
• In silico scores (ipSAE) predict wet-lab binding but aren't the same as confirmed wet-lab results — the cost and speed gains are for computational design, not validated drug candidates; a design binding well in silico is an early step, not a finished therapeutic.
• "Big" mode's largest-scale predictions (31,000-70,000+ tokens) are explicitly reported as inaccurate — the achievement is running inference at that scale on modest hardware, not correctly predicting structures that far beyond training context yet.
• The Life Sciences Verification Program is beta and gated — access requires application and verification, not open availability.

What this means for builders

The transferable lesson isn't really about biology specifically — it's about what happens when you point a capable agentic model at a well-defined performance bottleneck with a domain expert supervising rather than doing the engineering directly. Two Anthropic staff with zero kernel-engineering background supervised work that "normally takes an experienced team of engineers weeks," across more than 30 different model architectures, in under a month. That's the same shape of result explainx.ai has tracked in Claude's autonomous work on Fermat's Last Theorem and in NVIDIA's own accelerated inference runtime work — domain-specific performance engineering is increasingly something a supervised agent can do at a fraction of the specialist headcount it used to require.

Related on explainx.ai

• Claude designed working protein binders for 14 of 15 targets
• NVIDIA BioNeMo Inference Runtime: faster Boltz-2, OpenFold2, Protenix v2
• Claude Science: Anthropic's AI workbench for scientists
• Novo Nordisk and Claude Science for drug discovery
• Claude proved Fermat's Last Theorem in Lean
• Anthropic's embedded evaluators and pacing-the-frontier program
• AI drug discovery: clinical evidence and benchmarks
• Official source: Anthropic — How Claude is uplifting biomolecular modeling
• Official source: Anthropic — Introducing the Life Sciences Verification Program

This post is sourced to Anthropic's own September 17, 2026 research post and technical report. Speedup figures, cost comparisons, and the Life Sciences Verification Program's terms are Anthropic's self-reported claims; no independent third-party audit was available at time of writing.