Anthropic Enters Drug Development as OpenAI Offers US Government Equity

A week dominated by Anthropic’s scientific ambitions, OpenAI’s political manoeuvring, and a quiet but important new research benchmark — all while the cultural friction around AI-generated content continues to intensify.

The story of the week

Anthropic made two significant moves in quick succession. First, it launched Claude Science, a dedicated workbench designed to let researchers orchestrate experiments, pull together fragmented datasets, and generate figures — essentially doing for science what Claude Code already does for software engineering. Announced at a pharma-and-biotech-focused event, Claude Science can autonomously carry out meaningful research tasks from high-level instructions. Anthropic is framing this not as an incremental feature but as a push toward developing its own drug candidates, a move that would take the company from toolmaker to pharmaceutical competitor — a substantial escalation in ambition and, inevitably, in regulatory exposure.

Second, after weeks of negotiation with Washington, Anthropic got clearance to restore Claude Fable 5, the model that had been pulled offline amid policy disagreements with the Trump administration. Access is being re-enabled across Claude’s own platform as well as AWS, Google Cloud, and Microsoft. The episode is a useful reminder that model availability is now partly a political variable, not just a technical one.

New models & capabilities

Google DeepMind released Nano Banana 2 Lite and Gemini Omni Flash, expanding its model line-up at the lighter, faster end of the spectrum — the part of the stack that matters most for on-device and latency-sensitive enterprise applications. Separately, DeepMind announced a research partnership with A24, the film studio, suggesting a serious push into creative-industry workflows beyond image generation.

Google also shipped a notable product update: NotebookLM can now generate 60-second vertical video clips summarising uploaded research — a TikTok-format digest of your own documents. Clever, though one wonders how many board-level briefings will now arrive as portrait-mode reels.

Meanwhile, the new Google Home speaker landed to a verdict that neatly captures the state of consumer AI hardware: good build, half-baked intelligence. Gemini is not yet ready to make the smart speaker compelling beyond its existing repertoire of timers and light switches.

Research worth knowing

OpenAI published GeneBench-Pro, a new benchmark that tests AI performance specifically on genomics, biology, and scientific research using complex, real-world datasets. This matters for the industry because existing benchmarks skew heavily toward coding and general reasoning; GeneBench-Pro forces models to prove themselves in messy, domain-specific territory where errors have material consequences. It also positions OpenAI as a standard-setter in life-sciences AI — conveniently timed against Anthropic’s Claude Science launch.

A separate and genuinely interesting piece from MIT Technology Review examined LLM groupthink: ask any major model for a random number between 1 and 10 and you will almost always get 7. The convergence across models is a symptom of training on overlapping data and similar optimisation targets. A startup profiled in the piece is attempting to inject genuine diversity of output — a problem that matters far more in business contexts (scenario planning, creative ideation, risk modelling) than most operators realise.

On the security side, researchers demonstrated that AI-powered browsers can be manipulated by feeding them trivially false premises — telling an LLM that 2+2=5 is sufficient to collapse its guardrails. A pointed illustration of why agentic browsing remains a liability rather than a feature for anything involving sensitive data.

CEO watch

Sam Altman floated giving the US government a 5 percent ownership stake in OpenAI — framed as a way to share the upside of AI with the public and smooth things over with the Trump administration. The proposal, reported by the Financial Times, is nakedly transactional: equity in exchange for regulatory goodwill. Whether or not it materialises, the signal is clear — the largest AI labs now treat government relations as a capital-structure question, not merely a compliance one.

OpenAI also released new adoption data showing ChatGPT usage expanding across regions and languages, with users exploring deeper capabilities over time. And, in a move aimed squarely at European policymakers, OpenAI published a report mapping how AI could reshape jobs across the EU, identifying which occupations face automation, growth, or workflow change.

What it means for European operators

Three things to act on this week:

1. Claude Science changes the vendor calculus for R&D-heavy businesses. If you operate in pharma, chemicals, materials, or biotech, Anthropic is now competing for a slot that was previously occupied by specialised scientific-computing vendors. Evaluate it seriously, but watch the data-residency terms: Anthropic’s infrastructure remains overwhelmingly US-based, and European health-data rules apply the moment patient or clinical-trial data enters the picture.

2. OpenAI’s EU jobs report deserves a sceptical read — and a strategic one. The mapping exercise is useful input for workforce planning, but it is also lobbying: OpenAI wants to shape the narrative around AI and employment before European regulators do. Use the data; do not adopt the framing uncritically.

3. The LLM groupthink problem is your problem. If you are using models for any form of analysis that depends on diversity of perspective — risk assessment, market research, scenario generation — the convergence issue is not theoretical. Multi-model architectures and explicit temperature/sampling strategies are no longer optional; they are basic hygiene for decision-support systems. The fact that every frontier model gravitates to the same “random” number should give any operator pause before trusting a single model’s output on anything consequential.