Can AI Think Like a Scientist? OpenAI’s New Life Science Benchmarks Put It to the Test

FEATURED INSIGHT💡

LifeSciBench Tests Whether AI Can Be a Useful Scientific Collaborator

Most biology benchmarks reward models for retrieving the correct fact or producing a clean prediction. Real research is far less cooperative. Scientists reconcile conflicting studies, inspect figures and sequence files, question whether assays are measuring the right thing, evaluate translational risk, and make decisions with incomplete evidence. LifeSciBench was designed around that reality, with 750 expert-authored tasks spanning seven scientific workflows and seven biological domains. Its construction involved 173 scientist contributors, 453 independent reviewers, 1,062 supporting artifacts, and more than 19,000 task-specific grading criteria.

The benchmark evaluates free-response work that might realistically be assigned to an experienced scientific collaborator. One task, for example, asks a model to pressure-test the evidence behind a Duchenne muscular dystrophy gene therapy program. A strong response must catch problems involving antibody specificity, protein quantification, biopsy design, surrogate endpoints, external control groups, durability, cardiac safety, and patient selection. Reaching the broad conclusion is not enough. The reasoning must be scientifically defensible and useful to someone deciding what experiment, analysis, or regulatory step should come next.

The results show meaningful progress alongside some sharp limitations. GPT-Rosalind achieved a 36.1% exact pass rate, up from 25.7% for GPT-5.5, with its strongest performance in scientific communication and translating research evidence into clinical implications. Performance dropped when tasks required exact calculations, experimental design, complex artifacts, or operational decisions. GPT-Rosalind passed 45.1% of text-only tasks but only 28.1% of tasks involving artifacts or URLs. For research teams, the immediate opportunity may be carefully bounded assistance: evidence synthesis, structured critique, scientific writing, and decision support with expert review still firmly in the loop.

Cut Lead Review From Hours To Minutes

Sign up for a free trial of Attio, the agentic CRM.

Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.

Review your pipeline in Claude, synced live from Attio via MCP.

That's it.

Try Attio Now

ON THE HORIZON 🌅

GeneBench-Pro and the Search for “Research Taste”

LifeSciBench asks whether AI can contribute across life science workflows. GeneBench-Pro goes after an even harder capability: knowing how to approach a messy computational biology problem in the first place. Its creators call this research taste, meaning the chain of judgments that determines which questions the data can support, which diagnostics matter, when assumptions should change, and when a result is reliable enough to inform a real decision.

The benchmark contains 129 problems across 10 domains and 21 subdomains, including statistical genetics, regulatory omics, pharmacogenomics, cancer genomics, proteomics, and clinical diagnostics. Each agent receives an imperfect dataset, limited experimental context, and a specific target estimand. It must explore the data, detect irregularities, select an analytical method, revise its approach when necessary, and return a precise final result. The datasets are synthetically generated from known causal structures, allowing answers to be graded deterministically while still preserving the ambiguity and complications of real biological data.

GPT-5.6 Sol reached a 28.7% pass rate at maximum reasoning and 31.5% in Pro mode. That remains below one-third, but it represents a steep improvement from the original GeneBench, where the best frontier model scored below 5%. Expert reviewers estimated that a typical problem could require 20 to 40 hours of human work, compared with several dollars in model inference costs. Even partial reliability could make these agents valuable for exploring datasets, testing workflows, and accelerating hypothesis triage. The remaining failures are revealing: models frequently spot individual issues but struggle to connect them into a complete, decision-ready analysis. Closing that loop may become one of the most consequential frontiers in scientific AI.

LATEST IMPORTANT NEWS 📰

OpenAI Shuts Down Atlas and Moves the Browser Into ChatGPT

OpenAI is retiring Atlas, the standalone AI browser it launched last year, while moving many of its agentic features into products people already use. A new Chrome extension will let ChatGPT understand the current webpage, answer questions, summarize content, and begin longer tasks without leaving the browser. The ChatGPT desktop app is also gaining a fuller browser that can visit websites, log into accounts, download files, and interact with pages, supported by a remote cloud browser for autonomous tasks. The strategy suggests that OpenAI sees browsing as a capability that should follow users across their existing workflow rather than requiring them to adopt an entirely new browser.

Google Will Disclose When Ads Are Made With AI

Google is adding a “How this ad was made” section to My Ad Center, where users can see whether an advertisement was created or edited with AI. Ads made through Google’s own generative AI tools will be labeled automatically, while advertisers using outside tools will generally need to disclose the use themselves. The feature will appear across Search, YouTube, and Discover. For brands and agencies, AI provenance is becoming part of the advertising workflow, which means creative teams may need clearer records of how assets were produced and stronger review processes for synthetic product imagery.

Prime Intellect Raises $130 Million to Build Enterprise-Owned AI Agents

Prime Intellect raised a $130 million Series A at a $1 billion valuation to help organizations train and operate specialized AI agents without relying entirely on frontier labs. Its platform combines compute access, reinforcement-learning infrastructure, and evaluation tools in a modular system that companies can adapt to their own workflows. Customers already include Ramp and Zapier, and the company says it has reached a $100 million annualized revenue run rate. The larger trend is enterprise ownership: organizations increasingly want models trained around proprietary knowledge, predictable costs, and infrastructure they can control rather than a critical workflow that depends on a third-party model provider’s roadmap.

Hampton took $440K in planned hires off the calendar

Hampton co-founder Joe Speiser had three roles budgeted: a data engineer, an ops manager, a PM. $440K. He installed Viktor on April 12. Forty-four days later, none are on the calendar, and 18 of his team work with Viktor daily. His VP: we are editors now, not creators.

Get Started for Free

FOR THE TECHNICALLY INCLINED 🛠️

SWE-1.7 Shows How Far Post-Training Can Still Go

Cognition’s new SWE-1.7 coding model scored 42.3% on FrontierCode 1.1 Main, nearly matching GPT-5.5 at 43.0% and trailing Claude Opus 4.8 at 46.5%, while operating at substantially lower cost. The model was trained from a Kimi K2.7 base that had already received extensive reinforcement-learning post-training. Cognition’s additional gains challenge the idea that heavily post-trained models quickly hit a capability ceiling. SWE-1.7 is optimized for long-running software engineering work and is available inside Devin through Cerebras at up to 1,000 tokens per second.

The training system may be as important as the model itself. Cognition used top-p sampling replay to preserve exploration during long reinforcement-learning runs, distributed rollout generation across four data centers on three continents, and transferred compressed weight updates that were more than 99% smaller than sending the full model. SWE-1.7 also learned to summarize its own working state and continue from that summary, allowing some training trajectories to run for as long as six hours. The resulting agent explores codebases more carefully, probes ambiguous behavior with experiments, and catches more edge cases, although that extra diligence can also cause it to modify more files than a task strictly requires.

AI TOOL OF THE DAY 🚀

Mercury Command is an AI financial operator built into Mercury that can analyze spending, send invoices and payments, issue cards, categorize transactions, and create transfer rules from natural-language requests, with every action reviewed and approved before execution.

10x the context. Half the time.

Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.

Try Wispr Flow free