The Future of Life Institute published its 2026 AI Safety Index this week.
It is not a think piece. It is a scorecard. Independent reviewers assessed each frontier lab across risk management, transparency, and whether the company's public safety commitments match its actual behaviour. Then they assigned letter grades — the same way a teacher would.
Nobody got an A. Nobody got a B.
Anthropic
Highest grade awarded to any lab
C+
OpenAI
The most used AI product on earth
C
Google DeepMind
Deepest research bench in the industry
C
Meta
Largest open-weight model distributor
D+
xAI · DeepSeek · Mistral
Effectively failed the assessment
F
Read that table again slowly. The most capable technology of this decade, the one being deployed into hospitals, courts, classrooms, and half the software you use is being built by organisations whose best safety grade is a C+.
Not one of them would pass as an honours student. Most would be having a conversation with their parents.
Now put the grades next to the money
Here is the number that turns a report card into a story.
US startups raised roughly four hundred and twelve billion dollars in the first half of this year. Eighty-six percent of it, around three hundred and fifty-six billion, went to AI companies.
Nearly nine of every ten venture dollars invested in America this year chased the same industry that just posted a C average on its own safety promises.
US startup funding, H1
412.7B
Dollars, first half of 2026
Share that went to AI
86%
Most concentrated on record
Left for everything else
14%
Biotech, consumer, fintech, climate
Best safety grade
C+
Across the entire industry
US venture funding, first half of 2026, where the money went
The part that should worry founders more
In past booms, a record funding year meant lots of different companies got money. This year, a record year means a handful of AI giants absorbed almost everything while every other founder watched the room empty out. If you are building outside AI, you are not imagining the silence, you are competing for fourteen percent of a market that used to be a hundred.
An industry that cannot pass its own safety exam is being handed the largest concentration of capital in the history of American venture funding. Both things are true at the same time, and neither is a prediction.
Presented by Atoms.dev
Tell Atoms your idea. A team of AI employees - PM, engineer, SEO specialist, data analyst that builds it, deploys it, and helps you acquire customers. 165,000 GitHub stars. #1 Product Hunt of the Week. Trusted by Amazon, Microsoft, and NVIDIA.
What a “C” actually means for the thing you built
Grades are abstract. Consequences are not. The index scored labs on three things and each one maps directly onto a risk sitting inside your product right now.
Risk management: the outage you cannot control
A mediocre risk grade means the lab you build on may pull, gate, or change a model with little warning. We have watched it happen twice this year already: a frontier model was suspended globally over export controls, and another was re-priced from included to metered overnight.
If your product breaks when your model does, the lab's risk grade is now your risk grade. Know which model you depend on, and know what your fallback is.
Transparency: the thing you cannot answer for
When a customer's legal team asks why your product recommended what it recommended, you inherit whatever your model provider is willing to disclose.
A C on transparency means the honest answer is often that nobody outside the lab knows. That is survivable in a side project. It is not survivable in an enterprise procurement conversation, and enterprise buyers have started asking.
Promises kept: the gap between the blog post and the behaviour
The index specifically measures whether public safety commitments match actual behaviour. Every lab scored poorly here.
The practical lesson for founders: do not build your roadmap on a lab's stated intentions. Build it on their shipped behaviour. Announcements are marketing. Changelogs are evidence.
This connects to last week. We covered Anthropic's letter to the Senate alleging the largest AI copying operation ever documented. The same lab that just earned the industry's highest safety grade is also the one arguing loudest that the rules need teeth. Read The Claude Heist in the archive →
🔮 The Bottom Line
The best safety grade in artificial intelligence this year was a C+.
Not from a critic. Not from a competitor. From an independent index measuring whether these companies do what they say they do.
Nine of every ten venture dollars in America went to them anyway. That is not hypocrisy, it is just what happens when the upside is large enough that nobody wants to read the report card.
You are building on top of these labs. You do not get to skip the reading.
📧 Forward this to 3 entrepreneur friends who need to see this opportunity
