How good is AI at real science?

A tougher test for scientific reasoning

The gap shows up in the scores.

OpenAI released FrontierScience, a new benchmark with over 700 expert-written questions across physics, chemistry, and biology. GPT-5.2 scored 77% on the Olympiad-style problems and 25% on the research-style tasks that use a 10-point rubric.

This confirms something scientists already feel at their desks late at night, staring at a bright monitor while a fan hums. The models are fast and reliable at tightly framed reasoning, but they still struggle when the task looks like real research with ambiguity, judgment, and multi-step planning that is not fully specified.

FrontierScience is useful because it avoids saturated multiple-choice tests and forces models to show their work, even if a model grades the output. The low research score also sets expectations, since this is assistive tech today, not an autonomous scientist.

If benchmarks keep getting closer to real work, what happens when models start missing fewer of these open-ended steps?

Google wants your day to start in Gmail

Google is back in your inbox.

The company is testing CC, an email-based productivity assistant from Google Labs powered by Gemini. CC sends a daily “Your Day Ahead” email that summarizes Gmail, Calendar, Drive, and tasks, and lets users reply by email to add to-dos, save notes, or search for information.

This feels like a conservative bet, and that is probably the point. Email is already where people check first thing in the morning, usually with a laptop open and coffee cooling nearby, so Google is meeting users where they already are instead of pushing another app.

The limitation is clear since CC only works for consumer accounts, not Workspace, and is gated to AI Pro and Ultra users in the U.S. and Canada. Still, it shows Google sees the inbox as a control surface for AI, not just a message queue.

If assistants live in email, do they quietly replace task managers, or just add another daily summary people stop opening?

Amazon tests the OpenAI waters

The number is doing the talking.

Reuters reports Amazon is in talks to invest about $10 billion in OpenAI, in a deal that could value the company above $500 billion. OpenAI is also preparing for an IPO that could reach a $1 trillion valuation, while Microsoft already owns a 27% stake and controls cloud resale rights.

This looks less like a surprise bet and more like a positioning move as compute supply tightens. Amazon wants closer access to frontier models and a reason to push its Trainium chips, which OpenAI is reportedly planning to use instead of relying only on Nvidia.

If this goes through, OpenAI ends up with two hyperscalers tied to its future, each with different incentives around cloud, chips, and enterprise customers. That setup gives OpenAI leverage, but it also raises the risk of subtle constraints showing up later in pricing, access, or integration.

When every cloud giant wants a piece of the same model, who actually gets to steer where it goes?

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.