Anthropic AI for Science: John Jumper, Claude Mythos 5 — ускорение drug design в 10 раз
30 июня 2026 — Anthropic проводит в Сан-Франциско The Briefing: AI for Science (глобальный стрим, 1 июля 01:00 МСК). Не рутинный product drop: лауреат Нобелевской премии по химии, соавтор AlphaFold John Jumper переходит в Anthropic; Claude Mythos 5 даёт ~10× на ключевых шагах drug design; Novo Nordisk режет время на CSR на 90%. CEO Novartis, BMS и Genentech — на сцене.
Для life sciences, pharma decision-makers и AI-инженеров разбираем три блока: контекст ивента, почему Jumper ушёл с DeepMind на пике; 18-месячный roadmap, Claude for Life Sciences и hard benchmarks Mythos 5; экспортный контроль (12/26 июня), логика покупки Coefficient Bio и actionable 6-step playbook. Cutoff: 30.06.2026, до окончания live.
01 Anthropic AI for Science 30 июня 2026: что за ивент
Anthropic объявляет войну life sciences — вторая вертикаль после Claude Code. Pharma R&D — один из немногих доменов, где LLM реально может заменить expert labor.
| Параметр | Значение |
|---|---|
| Event | The Briefing: AI for Science |
| Datetime | 30 июня 2026, 10:00 AM PST (1 июля 01:00 МСК) |
| Format | San Francisco onsite + global livestream |
| Host | Anthropic (parent Claude) |
| Agenda | Life sciences vision, demos, enterprise case studies |
Speaker lineup (выборка)
| Имя | Роль |
|---|---|
| Vas Narasimhan | CEO Novartis, board member Anthropic |
| Chris Boerner, PhD | CEO Bristol Myers Squibb |
| Aviv Regev | EVP R&D, Chief Scientific Officer Genentech |
| Lotte Bjerre Knudsen | ex-CSO Novo Nordisk, профессор DMSc |
| Eric Kauderer-Abrams | Head of Life Sciences, Anthropic |
| Jonah Cool | Head of Life Sciences Partnerships, Anthropic |
| Matthew Herper | Senior pharma reporter, STAT News (moderator) |
- Pain #1 — R&D cycle: 12–15 лет, >$2.6B per drug; ~10% clinical candidates доходят до approval.
- Pain #2 — compliance docs: CSR и regulatory filings — месяцы ручной работы, bottleneck на filing.
- Pain #3 — tool sprawl: Benchling, PubMed, 10x Genomics в silos — context-switch tax на учёных.
- Pain #4 — model access: Mythos 5 частично недоступен из-за US export controls — compliance uncertainty для global teams.
02 John Jumper: AlphaFold, Нобель и переход в Anthropic
John Michael Jumper, р. 1985, Little Rock, Arkansas. Track: Vanderbilt math/physics (2007) → Cambridge physics, Marshall Scholar (2008) → Chicago theoretical chemistry PhD (2017). Через 6 месяцев после защиты — Google DeepMind, secret project AlphaFold.
AlphaFold: 50-летняя open problem закрыта
Protein folding: predict 3D structure from amino acid sequence. CASP14 (2020): команда Jumper + Demis Hassabis обходит всех конкурентов по accuracy.
- 214M+ predicted protein structures (~1M species)
- 190 countries, 2M+ researchers
- Impact: cancer therapeutics, drug discovery, molecular biology
2024: Jumper + Hassabis — Нобелевская премия по химии (вторая половина — David Baker, UW). Jumper в 39 лет — самый молодой chemistry laureate за 70+ лет.
Анонс 19 июня 2026
«После почти девяти лет я ухожу из Google DeepMind и присоединяюсь к Anthropic.» — John Jumper, X
Hassabis: «AlphaFold changed the world and proved what AI can do in science and medicine.» Анонс за 11 дней до сегодняшнего ивента — явный strategic narrative. Official title у Anthropic пока TBD; вероятно lead fundamental bioscience AI, возможен «ClaudeFold».
Reality check: реплицировать AlphaFold в Anthropic — underdetermined. Успех = years of DeepMind infra + top biology partnerships + verifiable CASP benchmark. Anthropic — LLM-first commercial shop, pivoting to scientific AI. Biological knowledge valuable, но product breakthroughs требуют org-wide execution.
03 18-месячный roadmap: Claude for Life Sciences stack
Не greenfield launch — кульминация systematic build. Апрель 2026: Series H $65B, valuation $965B; life sciences — pillar IPO story.
| Date | Event |
|---|---|
| Oct 2025 | Claude for Life Sciences GA — Benchling, 10x Genomics, PubMed connectors |
| Feb 2026 | Partnerships: Allen Institute, HHMI (Janelia) |
| Apr 2026 | Acquire Coefficient Bio, ~$400M all-stock |
| May 2026 | Andrej Karpathy joins pre-training team |
| Jun 9, 2026 | Claude Fable 5 + Mythos 5 ship — life sciences leap |
| Jun 12, 2026 | US gov forces takedown Fable 5 & Mythos 5 (export control) |
| Jun 19, 2026 | John Jumper leaves DeepMind → Anthropic |
| Jun 26, 2026 | Partial restore: ~100 US critical infrastructure orgs get Mythos 5 |
| Jun 30, 2026 | AI for Science launch event |
Claude for Life Sciences = vertical pharma solution on Claude Enterprise. Core: MCP connectors + Agent Skills across full pipeline.
| Platform | Use case |
|---|---|
| Benchling | ELN/LIMS — SOP, consent forms |
| 10x Genomics | Single-cell seq, spatial transcriptomics |
| PubMed / bioRxiv / medRxiv | Literature + preprint analysis |
| Open Targets | Target ID and prioritization |
| Medidata / ClinicalTrials.gov | Clinical trial monitoring |
| Wiley Scholar Gateway / BioRender | Paper access, scientific figures |
Pipeline coverage: early discovery → preclinical (genomics, scRNA-seq QC, tox prediction) → clinical (protocols, enrollment) → regulatory (filings, gap analysis, FDA responses).
04 Claude Mythos 5: drug design scores и pharma deployments
Strongest Anthropic model for life sciences. Internal evals with protein design + bioinformatics tooling, zero human assist:
| Metric | Result |
|---|---|
| Speedup | ~10× on critical design steps |
| Hit rate | 9/14 protein targets (64%) → strong candidate compounds |
| Autonomy | binding site → tool pick → design run → self-recovery from failures |
| Target classes | immune checkpoints, growth factor pathways, neurodegeneration, muscle disease, complex structural targets |
| AAV (Dyno Therapeutics) | beats specialized protein LMs on capsid structure prediction |
| Hypothesis blind test | ~80% human reviewer preference vs Opus-gen |
| Lab validation | E. coli antibiotic target hypothesis — preliminary confirmation |
| Unsupervised week | 138 animal species, millions of cells → custom ML model 100× smaller than recent Science baseline, better perf |
Novo Nordisk NovoScribe: CSR -90%
Novo Nordisk (Ozempic) built NovoScribe on Amazon Bedrock + Claude, RAG + expert-validated templates.
«Claude cut our CSR writing time by 90% — docs go straight to human review and approval.» — Waheed Jowiya, Director Digital Strategy, Novo Nordisk
Scope expanding: device protocols, patient materials, CTD full-pipeline automation in exploration.
| Company | Use case |
|---|---|
| Novo Nordisk | NovoScribe CSR/CTD automation |
| Sanofi / AbbVie / AstraZeneca / Genmab | Enterprise Claude for Life Sciences rollout |
| Bristol Myers Squibb | CEO stage + power user |
| Komodo Health | Healthcare data analytics |
| Axiom | Claude Code + MCP tox prediction |
05 Coefficient Bio acquisition и competitive moat
April 2026: Anthropic acquires Coefficient Bio for ~$400M all-stock — stealth biotech, <10 FTE. Founders Samuel Stanton + Nathan C. Frey from Genentech Prescient Design computational drug discovery. Thesis: «ASI for Science». Investor Dimension: 38,513% IRR. Team under Eric Kauderer-Abrams — protein design + biomacromolecule modeling = bridge from assistant to AI drug discovery engine.
AI reshapes drug dev cost curve
- Target ID: millions of papers → signal extraction, months → hours
- Compound design: wet-lab screening → compute simulation, order-of-magnitude speedup
- Clinical docs: FDA-compliant draft generation, multi-x compliance efficiency
Anthropic vs OpenAI / Google DeepMind — three edges:
- Safety-first culture: Constitutional AI → regulatory trust in pharma
- Vertical integration: connectors + Coefficient Bio + Jumper science credibility
- Enterprise lock-in: Novartis, BMS, Genentech, Novo Nordisk reference customers
- Citable · R&D cycle: 12–15 years avg, >$2.6B (2024 data)
- Citable · clinical success: ~10% of trial candidates reach market
- Citable · AlphaFold scale: 214M+ structures, 190 countries, 2M+ researchers
- Citable · Mythos 5: 10×, 9/14 targets, 80% hypothesis preference rate
- Citable · Coefficient Bio: $400M deal, 38,513% IRR (Dimension)
06 Export control, outlook и FAQ
Mythos 5 — strongest life sciences weapon, under US government control:
| Date | Action |
|---|---|
| Jun 12 | Forced takedown Fable 5 & Mythos 5; non-US citizens blocked |
| Jun 26 | Partial restore: ~100 US critical infrastructure orgs |
| Fable 5 | Full restore still in negotiation |
Material compliance risk for global pharma. Deep dive: Fable 5 ban and alternatives.
Key theses
- Life sciences = next AI battlefield after Claude Code
- Jumper hire = bet on fundamental science AI, not just report writing
- China / non-US outlook: Mythos 5 access path unclear; Baidu, Alibaba Cloud, Zhipu as domestic alternatives
Watch today's event for
- Jumper appearance + official role
- Mythos 5 biology open program or trusted access expansion
- New Claude for Life Sciences connectors / Agent Skills
- Fable 5 restore timeline
- International access path for non-US researchers
FAQ
Q: Насколько Mythos 5 быстрее humans в drug design?
A: ~10× on critical steps; 9/14 protein targets produced candidates with zero human assist.
Q: Можно ли использовать Mythos 5 для biology research сейчас?
A: As of 2026-06-30: Mythos 5 only for ~100 vetted US orgs; Fable 5 paused since Jun 12 export control.
Q: Почему Coefficient Bio стоит $400M?
A: Genentech computational drug discovery talent; protein design + biomacromolecule modeling = critical jump to discovery engine.
Sources: Claude for Life Sciences, Fable 5 & Mythos 5, Novo Nordisk case, Jumper Nobel lecture.
07 Six-step playbook для life sciences teams
- Inventory pipeline pain points: literature/target ID, compound design, CSR/CTD writing, trial monitoring — prioritize highest ROI.
- Audit connector coverage: Benchling, PubMed, 10x Genomics, Medidata in your stack? Gap-fill via API/MCP bridges.
- Clarify Mythos 5 / Fable 5 access: US entities — Project Glasswing or critical infrastructure whitelist; non-US — Opus 4.8, GPT-5.6, or compliant domestic models.
- Clone NovoScribe pattern: expert-validated templates + RAG + Claude/Bedrock backend; CSR pilot before CTD rollout.
- Human-in-the-loop review: no AI output into regulatory filing without domain expert sign-off; full audit logs for FDA/EMA traceability.
- Monitor post-event announcements: Jumper role, Mythos 5 open plan, Fable 5 timeline — any shift reshapes tool selection.
08 Итог: life sciences infra needs a stable base layer
Today's AI for Science event = public capstone of 18-month roadmap. Claude for Life Sciences infra, Coefficient Bio acquisition, Karpathy pre-training, John Jumper arrival — coherent bet that AI will compress discovery timelines at clinical scale. AlphaFold-level breakthrough TBD, but the puzzle is more complete than most observers assume.
Teams running Claude Code, self-hosted MCP connectors, local bioinformatics pipelines or 7×24 agent automation hit three SaaS API failure modes: export control can kill top-model access overnight, long jobs on shared cloud get throttled/preempted, cross-border compliance audits don't work in third-party environments. For production-grade agent + life sciences workloads: JEXCLOUD multi-region bare-metal Mac — dedicated Apple Silicon, 7×24 uptime, monthly scale, 120s provisioning, persistent MCP servers, local RAG indexes, data isolation. Pricing: JEXCLOUD pricing page.