The directions that can currently be confirmed are: a research workbench that contains no new model (Claude Science), an open interface for laboratory hardware with named vendors and labs already connected (MHS), a protein-design campaign executed by outside wet labs, two mathematical results (a formalization of Fermat's Last Theorem and an improved Riemann-related bound), a large access programme for scientists, and a wet lab of its own, which as of September 23 has reported its first finding. No domain models appear in the public record: almost everything here is the general Claude model plus a layer that makes its output checkable, and that layer now includes, for the first time, a bench result that was reviewed and tested by Anthropic's own scientists rather than an outside lab.
Scope note: this page tracks the 2026 public record only: official releases, papers, third-party reporting and named collaborator disclosures. No inside information. It records how far public material goes, not how well the organization is doing.
Overview
| Direction | What can be confirmed | Latest evidence |
|---|
| 1. Research workbench | Claude Science is not a new model; it is an integration layer over 60+ skills and data sources | Released 2026-06-30 |
| 2. Instrument interface | MHS has named vendors and test labs; CMU cut weeks of setup to about eight hours | Research preview 2026-08-27 |
| 3. Protein design | Binders against 14 of 15 targets, 22.6–35.1% hit rates, executed outside Anthropic | Published 2026-08-18 |
| 4. Formal mathematics | 11 days, 13 million lines of Lean, 29,500 intermediate theorems, reviewed by an outside mathematician | Published 2026-09-04 |
| 5. New mathematics | Critical-line zero fraction raised from 41.6% to 67.2%; the model is unreleased | Published 2026-08-10 |
| 6. Access programme | 10,000 seats; up to 50 projects at $30,000 of credits each | From 2026-07, expanded 2026-08-28 |
| 7. A wet lab of its own | First result published: agents flagged an uncharacterized DNA repeat array; lab experiments confirmed it is expressed as RNA. Function unknown; size and staffing still unpublished | Anthropic, 2026-09-23 |
1. Research workbench: Claude Science
What can be confirmed: released June 30, 2026 as "an AI workbench for scientists" that "integrates the tools and packages that researchers most commonly use, produces auditable artifacts, and provides flexible access to computing resources."
What is in it: more than 60 curated skills and connectors; preconfigured for genomics, single-cell, proteomics, structural biology and cheminformatics; NVIDIA's BioNeMo Agent Toolkit (Evo 2, Boltz-2, OpenFold3); scientific databases including UniProt, PDB, Ensembl, Reactome, ClinVar, ChEMBL and GEO; PubMed, Jupyter and R; Modal for compute on demand; SSH/HPC cluster access and local macOS/Linux deployment. Named early users include Manifold Bio, Allen Institute neuroscientist Jérôme Lecoq and UCSF Brain Tumor Center epidemiologist Stephen Francis.
The decisive detail: it is not a new model. Anthropic's own wording is that Claude Science is "not a new AI model and not a more capable model for biology. It runs the same Claude models already available to everyone today." The difference is at the workflow layer, not the model layer.
Boundary: the announcement states no limitations at all. Third-party reporting supplies one: the built-in citation and calculation check is performed by a reviewer agent, and "it's still the same underlying model checking itself, not an independent source of truth." Confidence: medium (official release plus named users; no external evaluation).
Open leads: whether any independently reproducible research result has come out of the workbench; how the 60-plus skills are actually used.
2. Instrument interface: Model Hardware Standard
What can be confirmed: a research preview opened August 27, 2026. MHS is a shared specification that lets agents safely operate physical devices in research labs and advanced manufacturing lines, covering instrument control, state, measurements and errors. It was developed with HHMI Janelia.
Who is already connected, by name: on the vendor side, AWS (through its Strands Robots package), Automata, Danaher, Doosan Robotics, MBF Bioscience, QIAGEN, Tecan and Universal Robots; on the open-source side, Hugging Face's LeRobot project. Test sites include Genentech, Carnegie Mellon University, QuEra Computing, Tetsuwan Scientific and University of Washington labs.
One checkable number: in early testing at Carnegie Mellon, drivers and orchestration for a hardware configuration that normally takes several weeks of vendor setup were completed in about eight hours. Public demonstrations include an autonomous rerun of a dose-response experiment, collision-free coordination between lab robots, and AI-assisted quantum-laser stabilization.
Boundary: this is a research preview, not a released standard, and "vendors are building support into their products" is not the same as "in routine production use." Confidence: medium-high (named vendors and named labs, with a checkable setup-time figure).
Open leads: which of those integrations are pilots and which are routine; how the specification will be governed.
3. Protein design: 14 of 15, run by outside labs
What can be confirmed: published August 18, 2026. Fifteen targets (sixteen were selected; one was dropped for inconclusive data), 30 designs requested per target, 1,320 designs generated, of which 354 binders were synthesised and tested. The models used were Claude Mythos Preview and Claude Opus 4.8.
Who executed it: Adaptyv Bio and Twist Bioscience "independently produced and tested Claude's designs in the lab": everything past the design step happened outside Anthropic.
Results: binders were obtained against 14 of the 15 targets. Hit rates were 26.7% (Mythos Preview, multi-target), 22.6% (Opus 4.8, multi-target) and 35.1% (Mythos Preview, single-target), against the stated field baseline that "10 to 15% is typical in protein design campaigns today." Against RBX1, Mythos Preview in single-target mode reached a 40% hit rate versus 3.7% among competition participants, and its top design outperformed the winning entry among 245 designs submitted.
Stated limitations: "we intend to follow our experiments with more extensive characterization to confirm our hit rates and affinity measurements," and the write-up names targets Claude struggled against, including BBF-14 and maltose-binding protein.
Confidence: high (external execution, a stated baseline, and named failures). This is the strongest evidence on the page: the work past the design step was not done in-house, and the targets that failed are written down.
What can be confirmed: published September 4, 2026. Working largely autonomously for 11 days, Claude produced an end-to-end, computer-checked formalization of Fermat's Last Theorem, writing 13 million lines of Lean and proving 29,500 intermediate theorems used in the final proof. It is the largest Lean proof written to date.
External review: Kevin Buzzard of Imperial College London reviewed it and said the autoformalization "proves Fermat's Last Theorem with no assumptions other than the axioms of mathematics," adding that "if the automatic formalization of FLT is possible now, then we have taken a big step towards automatic formalization of the modern mathematical literature."
Boundary: this is not new mathematics. Claude did not discover the proof and did not replace Wiles's 1995 work; Lean checks whether each step follows from definitions, axioms and earlier theorems. Mathematicians had expected the formalization to take years and it took 11 days. What changed is the cost, not the conclusion. Confidence: high (public artifact plus review by a named outside mathematician).
What can be confirmed: published August 10, 2026. An unreleased Claude research model raised the known lower bound for the fraction of Riemann zeta zeros on the critical line from 41.6% to 67.2%, with a formally checkable proof.
Boundary: it does not, and does not claim to, resolve the Riemann hypothesis. The model is not publicly available, so outsiders cannot reproduce how the result was produced, only check the proof it emitted. Confidence: medium (the conclusion is checkable, the process is not reproducible).
6. Access: seats and grants
What can be confirmed: the AI for Science programme gives free credits to high-impact research projects. In mid-2026 it expanded to support up to 50 Claude Science projects at up to $30,000 of credits each (plus up to $2,000 of Modal compute), with applications closing July 15, decisions by July 31, and projects running September 1 to December 1. From August 28, 2026, a Claude Team plan for scientists offers 10,000 seats across mathematics, chemistry, physics and other fields, free or discounted for a year.
Boundary: this is input, not output. The public record contains no results from those projects. Confidence: high (official release, explicit terms).
Why it is tracked: frontier-model access is becoming research infrastructure, which brings the allocation questions that compute, data and instruments already have: who gets a seat, on what criteria, judged by whom.
7. A wet lab of its own: the physical layer
What can be confirmed: on September 18, 2026, Reuters reported that Anthropic has built a wet lab in the San Francisco Bay Area, a place to run physical biology experiments rather than simulate them. Eric Kauderer-Abrams, the company's head of life sciences, confirmed its existence; a spokesperson said it is "not for drug discovery specifically" and declined to elaborate. Kauderer-Abrams: "We believe that to do biology, the final test is still and will be for a while in real lab work." He also said life sciences is now among Anthropic's largest areas by headcount and resources. The company has been recruiting a procurement and operations lead, and a posting asks for expertise in "protein and nucleic acid characterization". Reuters reported that Anthropic wants to push how Claude can direct robotic units to carry out experiments with limited human intervention; asked about that, Kauderer-Abrams said "We're in the very early innings of using AI to automate the execution of lab work."
First result, published September 23, 2026: Anthropic's own account says the lab runs at BSL-1 and BSL-2, handles no pathogens that infect humans, and that "all of the lab work is performed by human scientists." It sits inside Anthropic's life sciences organisation alongside drug-discovery teams and the group that trains Claude on biology and chemistry. A typical workflow has Claude survey a protein family, read the literature, reproduce established results to check its own methods, then search for family members or genomic neighbours that fit no described system. Run this way with a harness coordinating many parallel sessions, agents gathered over 200,000 reverse-transcriptase (RT) family members, narrowed those to 3,500 candidate systems, and analysed the 20 most compelling in depth. One agent, reading raw DNA next to an unusual RT, flagged a tandem repeat array no one had described: "I can see by eye a tandem repeat array … that's a CRISPR-like … repeat array?!" Anthropic scientists reviewed the report and tested the candidate on the bench. The system, named ART (array-associated reverse transcriptase), occurs mainly in bacteriophages; its repeat layout structurally resembles a CRISPR array, and a first wet-lab experiment confirmed the array is expressed as a set of distinct short RNAs. What ART actually does is still unknown, and further experiments are underway. A preprint accompanies the announcement.
Boundary: biosafety level and the fact of a first experiment are now public, closing two of the four gaps this page flagged before September 23, but size, staffing and opening date remain unpublished, and the finding itself stops at structural resemblance and a confirmed expression pattern: no function has been demonstrated, and this is a preprint, not yet peer reviewed, reported by Anthropic about Anthropic's own lab. Anthropic told Reuters it still sends part of its biological work to outside partners, so the lab is an addition, not a replacement. Confidence: medium (official account and a linked preprint; no independent review or replication yet).
Why it is tracked: the six directions above reduce to one sentence: the general Claude model, plus a layer that makes its output checkable, where the checking is always executed by somebody else: Lean, outside wet labs, a partner's instruments. A bench of its own changes that shape, and now there is a first case to point to: an agent noticed an anomaly nobody had flagged, and a human-run experiment on Anthropic's own bench checked it. The other half still has to be said as plainly: "Anthropic has an autonomous biology lab" is an overstatement today. Claude proposed and helped interpret; humans reviewed the report, chose to test it, and ran the experiment. What changed on September 23 is that "very early innings" now has one named result attached to it, not just a stated intention.
What could not be found
- Any independently reproducible research result produced with Claude Science;
- Which MHS integrations are in routine use rather than pilots;
- The fuller characterization of the protein-design hits that Anthropic says is coming;
- Whether the research model behind the Riemann result will be released;
- Any output disclosure from the 10,000 seats or the 50 funded projects.
- The wet lab's size, staffing and opening date; what ART actually does; independent review of the ART preprint.
Method and scope
Scope: only work where AI is applied to scientific research. General model capability and consumer products are excluded.
Confidence is assigned claim by claim, not to a whole direction. The labels:
- High: a public artifact, execution by an outside institution, or review by a named expert;
- Medium: an official release or company-reported figure, inspectable in detail, unverified outside;
- Low: a single-source claim, a spoken statement, or a claim under public dispute with review incomplete.
Two operating rules: revisions are dated by version and are not treated as new projects; evidence that a project exists is not evidence that the technology has succeeded.
Primary sources
Official releases and research
Third-party reporting
