Every few weeks a chart makes the rounds with a line climbing off the top of the frame and the letters AGI in the caption. The replies split immediately. Half the thread insists we are two years from machine minds; the other half points out that the model still cannot reliably book a flight without inventing a confirmation number. Both reactions make sense. AGI milestone updates have become the oddest genre in tech journalism, part measurement, part marketing, part Rorschach test.
The useful question is narrower than “are we there yet.” It is this: which specific capability crossed a threshold this year, who verified it, and what did it cost to produce? That framing turns a shouting match into something you can actually track quarter to quarter.
What separates a real AGI milestone from a press release
Three kinds of evidence carry weight. Test scores are the loudest and the least durable. Behaviour on long tasks is the most predictive. And the ability of a system to improve the systems around it is the one that most researchers treat as the actual dividing line.
Benchmarks saturate, then get quietly replaced
MMLU was a decent filter in 2021. By 2024 frontier models were scoring in the high 80s and it had become a formality. ARC-AGI-2 launched in early 2025 as a deliberate attempt to build something resistant to that fate; top models scored in the low single digits on release. Within months, reasoning models had pushed well past 20 percent on the same set. That is a real jump, not a curve-fit. It is also worth reading the fine print, because private holdout sets and public leaderboards tell different stories.
Task horizons beat test scores
The most useful metric of the past two years came from METR, a small evaluation nonprofit. Instead of asking how many multiple-choice questions a model answers correctly, they measure the length of task, in human working time, that a model can complete about half the time. That horizon has roughly doubled every seven months since 2023, moving from tasks measured in seconds to tasks measured in hours. When people ask whether AGI milestone updates are accelerating, this is the chart to look at, not a bar graph of exam results.
Automated research is the threshold that matters
When a model starts producing genuine improvements to the process that trains models, the timeline compresses in ways that ordinary scaling does not explain. Google DeepMind’s AlphaEvolve found a faster algorithm for a piece of matrix multiplication that had resisted human optimisation for years, and recovered a small but real slice of Google’s global compute in the process. Agentic coding benchmarks tell a similar story, with leading systems now resolving the majority of real GitHub issues in curated test sets. If you want a single signal to watch, watch whether labs can still explain their gains through human effort alone.
The updates worth paying attention to right now
Strip out the announcement language and a handful of concrete shifts stand out from the past year:
- Longer autonomous runs. Agents that once collapsed after four or five steps now hold a plan together for dozens, including tool calls, retries and error recovery.
- Economic task evaluations. OpenAI’s GDPval compared model output against experienced professionals across 44 occupations, and blinded graders preferred the model’s work on a meaningful share of tasks. That is a different claim than passing an exam.
- Reasoning at inference time. Spending extra compute to think before answering moved hard maths and science results substantially, and made the cost per query a design choice rather than a fixed price.
- Context and memory. Million-token windows are now table stakes, though retrieval fidelity degrades long before the advertised limit.
- Vision-language-action models. Humanoid and warehouse robots improved faster than most forecasts expected, largely by borrowing the same architectures as chat models.
None of these is AGI on its own. Stacked together, they explain why serious forecasters moved their timelines inward without ever agreeing on a date.
Where the progress keeps stalling
The gaps are consistent and they are stubborn. Continual learning remains unsolved: a model fine-tuned on new material tends to lose sharpness elsewhere, so most systems still learn nothing from yesterday’s conversation. Reliability past the 95th percentile is another wall. A coding agent that succeeds four times in five sounds impressive until you need it to succeed in production every day for a month.
Embodiment lags the software by a wide margin, which matters because a large share of human competence is physical. And the science of evaluation is genuinely behind the science of capability. Contamination, prompt sensitivity, and benchmark gaming all distort the picture, which is why a healthy scepticism belongs in every reading of a new result. There is a good breakdown of where the line between evidence and narrative currently sits in this look at AGI breakthroughs and the hype around them.
The definition fight is not academic
How you define AGI now decides who pays whom. OpenAI’s original arrangement with Microsoft included a clause tying AGI to a profit threshold, and arguments about whether that line has been crossed are arguments about billions of dollars and control of the most expensive infrastructure project in commercial history. Courtroom testimony in the Musk–Altman dispute put a lot of that internal reasoning on the record, and the reporting on how the trial shaped OpenAI’s standing is a useful reminder that definitions drafted in 2019 are now load-bearing.
Regulators are converging on the same problem from the other direction. General-purpose AI obligations in the EU, transparency rules in California, and procurement standards inside governments all need a trigger point. Every time a lab publishes an AGI milestone update, someone in a policy office has to decide whether it changes a legal obligation. That is why the wording of these announcements shifted from “we have built AGI” to “we are preparing for AGI” over the past eighteen months.
How to read the next announcement cycle
Most of what lands in your feed will be a capability demo wrapped in a roadmap. Five questions sort the substance from the noise:
- What was the evaluation, and was it run by someone with an incentive to fail?
- What is the failure rate on the task, not the success rate on the best attempt?
- How long a task was it, measured in human hours?
- What does it cost per completion, including retries and human supervision?
- Does it keep working after the demo ends, on inputs the team did not choose?
Meanwhile, the tools most teams actually use sit a long way from any frontier benchmark, and they are where the practical returns are. A marketing group that needs a client deck by Friday gets more value from a twenty-minute workflow for building a ten-slide sales deck than from any change in a reasoning benchmark. That distance between the frontier and the shop floor is one of the more underreported facts about this whole period.
Keep a simple log if you want to stay grounded. For each claim you see, note the date, the metric, the evaluator, and the cost. Six months from now, that log will tell you more about the trajectory of AGI progress than any single headline, and it will protect you from the two failure modes that catch most readers: believing a demo is a deployment, and believing that a slow quarter means the curve has flattened. Neither has been true yet.