Anthropic just showed its real playbook: own the operating layer for an industry instead of competing purely on model capability. The same model, packaged as a vertical workbench, becomes a different product with different pricing power. Worth understanding regardless of your industry.
- Anthropic launched Claude Science, a workbench giving scientists one environment for computational research, built on the same Claude models already publicly available with no new capability and no gating.
- A main AI assistant acts as a project manager, connecting to 60+ scientific databases, creating sub-assistants to delegate work, and handing tasks to custom “expert” assistants researchers build themselves.
- A separate fact-checking AI verifies citations and calculations before publication, though it remains the same underlying model checking itself rather than an independent source of truth.
- OpenAI took the opposite approach with GPT-Rosalind in April: a specialized fine-tuned model gated behind enterprise qualification, limited to partners including Amgen, Moderna, and Novo Nordisk.
- Anthropic will fund up to 50 postdoctoral and graduate research projects with up to $30,000 in credits each, with applications open through July 15, 2026.
What Happened
Anthropic introduced Claude Science this week, an AI workbench giving scientists one environment for computational research instead of bouncing between databases, pipelines, and separate tools, TechCrunch reported. Anthropic was explicit that Claude Science is not a new model and not a more capable model for biology. It runs the same Claude models, including Opus 4.8, with no special access.
The system works through delegation. One main assistant acts as a project manager, connecting to more than 60 scientific databases with pre-built toolkits for genomics, protein structure, and chemistry. It can spin up sub-assistants to split up work, or hand tasks to a custom “expert” assistant a researcher has built for their own specific work.
A separate fact-checking AI verifies citations and calculations before anything goes toward publication, addressing the growing problem of fabricated citations in AI-assisted research writing, though Anthropic acknowledges this is still the same model checking itself.
Claude Science can run on a lab’s own infrastructure rather than sending data to Anthropic’s servers, and it generates reproducible figures alongside the exact code and environment used to create them. Early users cited by Anthropic include a Gladstone Institutes scientist who built a genome browser in days, and an Allen Institute neuroscientist who built a multi-agent computational review pipeline.
The launch follows a different path than OpenAI’s April release of GPT-Rosalind, a model specifically fine-tuned for biological reasoning and gated behind enterprise qualification, with early access limited to partners including Amgen, Moderna, Allen Institute, Thermo Fisher, and Novo Nordisk.
Google DeepMind is taking yet a third approach, owning foundational science models like AlphaFold and AlphaGenome directly and bundling them with 30+ life science databases through Gemini for Science.
Why It Matters
The contrast between Anthropic and OpenAI’s approaches is the real story here. Anthropic is betting that workflow ownership, not model specialization, is the more defensible moat. This mirrors how Claude Code became the operating layer for software development without requiring a coding-specific model. If Claude Science works the same way for scientific research, Anthropic gains a vertical foothold that does not depend on winning the underlying model capability race, which matters as the gap between frontier labs continues to narrow.
The fact-checker layer is a meaningful but partial solution. Having the same family of models check its own outputs catches certain classes of errors but is not an independent verification source, and Anthropic is transparent about that limitation.
Pharma organizations named as Anthropic customer case studies, including Novo Nordisk, also appear as OpenAI partners, which means most serious life sciences organizations are testing multiple AI vendors in parallel rather than committing to one. Relve, an AI trends intelligence platform, is tracking how the workflow-versus-model strategic split plays out as frontier labs compete for vertical industry adoption beyond software development.
Bottom Line
Watch which approach research institutions and pharma teams actually standardize on over the next two quarters: Anthropic’s open-access workflow layer, OpenAI’s gated specialized model, or DeepMind’s foundational-model-plus-database bundle. The winner reveals whether workflow integration or model specialization is the stronger moat for vertical AI adoption.
For SaaS founders building in any vertical, the comparison matters beyond life sciences. Anthropic’s bet, that owning the operating layer beats owning a specialized model, is a strategic question every applied AI company in a regulated or technical vertical is currently working through. Watch how Claude Science performs in production before assuming the workflow-first approach is the safer one to copy.
