The best grade on the 2026 AI Safety Index is a C+ — and that scares me

The short version

The Future of Life Institute’s 2026 AI Safety Index handed out its top grade to Anthropic, and that grade was only a C+. OpenAI and Google DeepMind landed at C, Meta at D+, and xAI, DeepSeek and Mistral effectively failed. When the valedictorian of the AI safety class is a solid “eh, it’s fine,” that’s not a curve problem. That’s a whole-classroom problem.

What exactly did the report say?

The Future of Life Institute grades the biggest AI labs on how seriously they treat safety: their risk assessments, their transparency, whether they keep the promises they made, how they handle existential-scale risk, that kind of thing. For 2026, the highest mark on the board went to Anthropic at C+. OpenAI and Google DeepMind came in at C. Meta got a D+. And xAI, DeepSeek and Mistral scored so low they’re basically walking out of the exam room with a failing slip.

I went and read the 2026 AI Safety Index report so you don’t have to sit through the whole thing, and the part that stuck with me wasn’t any single grade. It was the framing. The people running this thing aren’t anti-AI cranks. The Future of Life Institute is staffed by researchers who mostly want these companies to succeed at building useful models. And their blunt takeaway is that even the labs trying hardest are quietly loosening the commitments they made a year or two ago, right as their models get plugged into cybersecurity tools, healthcare workflows and autonomous agents that take actions on their own.

That combination is the whole story. Standards drifting down while capability and access drift up. It’s like watching a car speed up while the driver slowly takes their hands off the wheel to check their phone.

Why does a C+ actually keep me up at night?

Because of what a C+ means when the subject is safety instead of, say, a history quiz.

A C+ on a pop quiz means you’ll live. A C+ on the systems that write code, triage patients and run agents across the open internet means the top performer in the entire field is doing a mediocre job at the one thing that keeps those systems from causing real damage. And everyone below them is worse.

Here’s the specific thing I keep chewing on. Anthropic built its whole brand around safety. Constitutional AI, responsible scaling policies, the founders literally left OpenAI over safety disagreements. If the company that made safety its identity tops out at C+, that’s not a ceiling set by laziness. That’s a ceiling set by how genuinely hard this problem is, plus how little external pressure there is to push past “good enough.” Nobody’s failing Anthropic out of the program. Nobody can.

And the failing grades matter in a different way. xAI, DeepSeek and Mistral aren’t fringe players. They ship models that millions of people and thousands of companies use. A failing safety grade on a widely-deployed model isn’t a private embarrassment. It’s a risk that gets distributed to everyone downstream who builds on top of it.

What does “walking back commitments” even look like?

This is the phrase that should get more attention than the letter grades. Labs make public safety pledges when they need goodwill, funding rounds, or a friendly regulatory hearing. Then, months later, some of those pledges get quietly softened, delayed, or reinterpreted into something much smaller.

A commitment to third-party evaluation before release becomes an internal review. A promise to publish detailed model cards becomes a thinner marketing summary. A pledge to hold back a capability until safeguards are proven becomes shipping it and patching later. None of these are dramatic. Each one is defensible in a meeting. Add them all up across an industry racing each other, and you get a slow erosion that no single decision is responsible for.

That’s the part the Index captures that a headline can’t. Safety isn’t collapsing in one big scandal. It’s leaking at the seams while everyone insists the ship is fine.

Is this just doom-mongering from safety people?

Fair question, and I asked myself the same thing. It’d be easy to write off the Future of Life Institute as a group that’s professionally incentivized to sound alarmed.

But look at who the grades reward and punish. The Index gives its best marks to the labs actually doing the work — Anthropic, OpenAI, DeepMind get credit for having real risk frameworks. It’s not a blanket “AI bad” pamphlet. It’s a report card that hands out A-level praise to nobody because nobody earned it, and hands out failing grades to labs that genuinely skip the homework. That internal consistency is what makes me take it seriously instead of filing it under hype.

Also, the timing is the argument. If these models were still just chatbots spitting out essays, mediocre safety would be a smaller deal. But 2026 is the year agents got real. Models don’t just answer now. They act. They write and run code, book things, move through systems with permissions. Mediocre safety on a system that only talks is annoying. Mediocre safety on a system that does is a different category of problem.

What should you actually do about it?

If you’re a normal person or a small business using these tools, you’re not going to fix industry-wide safety standards. But you can be a slightly smarter buyer.

When you pick a model to build on, the safety grade is now a real signal, not a nice-to-have. If two models are roughly equal on price and quality and one comes from a lab that failed a credible independent safety review, that’s a tiebreaker. You’re inheriting whatever risk that lab shipped.

For anything that touches sensitive data, money, or actions taken without a human check, lean toward the labs that at least show up to be graded and score in the passing range. A C+ isn’t great, but it beats a company that couldn’t be bothered to submit real risk documentation.

And honestly, keep a human in the loop on agent workflows longer than feels necessary. The whole reason this Index is worrying is that autonomy is arriving faster than the safety practices around it. You don’t have to match that pace in your own setup.

Where does this leave us?

The uncomfortable truth is that the market isn’t punishing bad safety, so labs have little reason to chase an A. Users chase capability and price. Investors chase growth. Safety is a cost center until something goes badly wrong in public, and then it’s a crisis. The Index is basically a group of researchers standing outside that dynamic holding up a sign that says the smartest kids in class are coasting.

I don’t think this means anyone should panic or unplug. The tools are useful and I use them every day. But a C+ being the ceiling is a fact worth carrying around in the back of your head, especially every time a lab announces something can now act on its own.

FAQ

Who topped the 2026 AI Safety Index? Anthropic received the highest grade, a C+. OpenAI and Google DeepMind followed at C, Meta at D+, and xAI, DeepSeek and Mistral scored failing grades.

Who runs the AI Safety Index? The Future of Life Institute, an organization focused on reducing large-scale risks from powerful technology, produces the Index by grading major AI labs on safety practices, transparency and whether they honor prior commitments.

Why is a C+ considered alarming? Because it’s the best grade in the entire field. When the top-performing lab only reaches C+ on safety while models are being wired into cybersecurity, healthcare and autonomous agents, the whole industry’s floor is low.

Does a failing grade mean a model is dangerous to use? Not automatically, but it signals the lab skipped credible independent safety review. For anything touching sensitive data, money or automated actions, treat a failing grade as a real risk factor when choosing tools.

Is the Index biased against AI? It rewards labs that do serious safety work and penalizes those that don’t, which suggests a consistent standard rather than a blanket anti-AI stance.