Ask ten AI researchers to define AGI and you’ll get eleven answers. Ask them what actually changed in the last nine months, though, and the answers start to converge. The latest AGI developments aren’t one dramatic breakthrough. They’re a pile-up of smaller ones: models that hold a train of thought for hours, agents that file their own pull requests, and benchmark suites that keep getting quietly retired because the frontier has already walked past them.
That shift matters more than any single launch. The story has moved from “look what the model can do in one prompt” to “look what it can still be doing four hours later.”
Reasoning stopped being a feature and became the floor
Two years ago, a model that visibly worked through a problem before answering was a product differentiator. Today it’s the default setting on almost every flagship system from OpenAI, Anthropic, Google DeepMind and xAI. The interesting question is no longer whether a model reasons. It’s how long it can keep reasoning without drifting, looping, or quietly deciding it’s finished when it isn’t.
Long-horizon tasks are where the gains show up
Benchmarks that measure sustained work rather than single answers have moved fast. Scores on software-engineering evaluations like SWE-bench Verified climbed from the mid-40s to the high 70s in roughly eighteen months across the top labs. Agentic task suites that ask a model to book travel, reconcile an invoice, or refactor a legacy module over dozens of steps have gone from embarrassing to merely imperfect.
The practical consequence is that people are handing over bigger chunks of work. A year ago, asking a model to “fix the failing test suite” was a party trick. Now it’s a Tuesday.
The reliability gap is the honest bottleneck
Here’s the arithmetic that keeps teams up at night. If a task needs forty sequential steps and each step succeeds 95% of the time, the whole thing lands about 13% of the time. Push per-step reliability to 99% and you’re at roughly 67%. That compounding curve explains why demos look magical and production deployments look tired. It’s also why so much current research effort goes into verification, self-checking and recovery rather than raw capability.
Agents left the demo stage
Coding agents are now ordinary infrastructure at a lot of companies. They triage issues, draft migrations, and open pull requests that a human reviews. Browser agents handle procurement portals and internal tools that never had an API. The unglamorous work of making this safe is what’s actually happening: permission scopes, audit logs, sandboxing, and cost controls that stop an agent from cheerfully burning $400 of inference on a task a human would have finished in ten minutes.
What’s changed culturally is expectation. Teams now assume an agent will attempt the work and ask for a human when it’s stuck, rather than the reverse.
The frontier got crowded again
For a while it felt like a two-horse race. That’s over. xAI’s Grok 4.6 release pushed hard on the frontier with aggressive pricing and long context, and it forced rivals to respond on cost as much as capability. Google, meanwhile, has been rebuilding its position after a stretch where its research prestige outran its shipped products. The internal story of how Argon is meant to return Google to the frontier is worth reading if you want to understand why the next twelve months look different from the last twelve.
Competition at this layer tends to compress prices and widen access. Both are good for anyone building on top.
The safety conversation got harder and more concrete
The extinction debate used to live in op-eds. Now it lives in hiring. When senior safety researchers leave a leading lab, the disagreement stops being abstract and becomes a public argument about how much capability work is too much. The fallout from an Anthropic exit turned into exactly that kind of debate, and it exposed a genuine fault line: do you build the thing to study it safely, or do you slow down and risk someone less careful getting there first?
What’s changed on the technical side is that evaluation is finally getting teeth. Third-party red-teaming, dangerous-capability thresholds written into deployment policies, and interpretability work that can actually point at a circuit rather than a vibe. None of it is solved. All of it is more real than it was.
Some of the most interesting AGI work is happening in wet labs
Language models are converging on similar capabilities. Biology is not. That’s why the money is moving. Zuckerberg’s $500M push into AI for biology is one signal among several, and it points at a broader bet: that the fastest route to genuinely new scientific capability runs through protein design, virtual cell models and lab automation rather than another chatbot.
This is arguably the most concrete progress toward the “cure diseases” version of the AGI pitch. Models that propose candidate molecules and predict assay results are already shortening research cycles in ways a general-purpose assistant cannot.
Signals worth tracking over the next two quarters
- Task duration records. How many hours of uninterrupted useful work a model can sustain before a human has to intervene. This is the number that actually maps to economic value.
- Cost per completed task. Not per token. Watch whether agents get cheaper at finishing things, because that’s what determines adoption.
- Evaluation turnover. Every time a benchmark gets saturated and replaced, you learn roughly how far the frontier has moved.
- Safety commitments with consequences. Deployment policies are only meaningful if a lab has actually paused something because of them.
- Scientific output. Peer-reviewed results where an AI system contributed something non-trivial, not just a preprint with a model name in the acknowledgements.
What this means if you’re building something
Design for unreliable genius. Assume the model will be brilliant on the first attempt at a hard problem and sloppy on the fortieth step of a boring one. Structure work so that failures are cheap, reversible and visible, and put a human checkpoint where the compounding errors accumulate rather than at the end.
Resist the temptation to rebuild your product every time a new model ships. The capability curve is steep enough that anything tied too tightly to one vendor’s quirks will be obsolete in a quarter, but the underlying jobs to be done, verification, recovery, permissioning, and turning messy real-world context into something a model can act on, are stable and still underbuilt.
The other thing worth doing is measuring your own work. Public benchmarks are useful for orientation and nearly useless for decisions. A small internal eval set built from your actual failed tickets, your actual customer emails, your actual edge cases will tell you more about the latest AGI developments than any leaderboard, because it tells you what the current generation of models can do inside your specific mess.