Ask five people what OpenAI’s AGI capabilities actually amount to and you’ll get five different answers, each delivered with more confidence than the evidence supports. The models pass the bar exam and still can’t reliably book a plumber. They earn a gold-medal score at the International Mathematical Olympiad, then lose track of a file path two steps into a workflow. Sorting out what these systems genuinely do means separating demonstrated behaviour from the marketing, the leaderboards, and the ambient dread on social media.
Reasoning models reset the baseline
The first real shift came in September 2024 with o1, which traded fast responses for a hidden chain of thought that let the model work a problem before answering. The gains were not incremental. On competition maths, the jump from GPT-4o to o1 looked like the difference between a strong student and a national-team contender.
By December 2024, o3 was posting 87.5% on the semi-private ARC-AGI set, an eval built to resist memorisation, and clearing roughly a quarter of FrontierMath’s research-level problems when most models sat near zero. The caveat matters: that ARC score came at about $20 of compute per task, thousands of times what a human spends. In July 2025 an experimental OpenAI model solved five of six IMO problems inside the four-and-a-half-hour contest window and wrote the proofs in Lean, a formal language a checker could verify. Nobody had to take the model’s word for it.
GPT-5, released in August 2025, wrapped this into a router that sends easy queries to a fast model and hard ones to a slower reasoning mode. For users the change is subtle. For measurement it is not, because the same product now spans a wide range of effort levels and “what GPT-5 scores” depends entirely on how much thinking you allow. The research direction behind those reasoning models is worth understanding if you want to read the next announcement properly.
Agents are where the ambition shows
Reasoning made models better at thinking. Agents made them try to act. Deep Research, launched in February 2025, spends five to thirty minutes browsing dozens of sources and returns a cited report. Operator and later ChatGPT agent click through interfaces on your behalf. Codex takes a ticket, writes a patch, runs the tests, and opens a pull request.
The compounding-error problem
Agent demos flatter the technology because demos are short. Real work isn’t. A model that completes each step with 95% reliability finishes a twenty-step task about 36% of the time. Push to fifty steps and you’re down near 8%. This is why coding agents shine on contained tasks and stumble on sprawling refactors, and why “it can use a browser” and “it can run your procurement process” sit in very different credibility brackets.
- Works well: scoped research, test-driven code patches, document extraction, form filling on a handful of known sites.
- Works with supervision: multi-file coding changes, data cleaning pipelines, competitive analysis across a dozen sources.
- Still fragile: anything needing dozens of sequential decisions, recovery from unexpected errors, or judgement about when to stop.
Benchmarks versus actual work
Benchmark saturation is real. SWE-bench Verified, MMLU, AIME: models have chewed through them faster than eval writers can build replacements, and contamination worries never fully go away. OpenAI’s own attempt at a harder measure is GDPval, published in September 2025, with 1,320 tasks drawn from 44 occupations that pay well and judged blind by experienced professionals. GPT-5’s outputs were rated at least as good as human expert work in close to 40% of comparisons, strongest on writing, research synthesis and software documentation, weakest where tacit context or physical judgement was required.
Forty percent is a strange number to hold in your head. It isn’t a rounding error and it isn’t a replacement. It describes a system that produces genuinely useful first drafts across a wide band of white-collar work while still needing a competent human to catch the misses.
The gaps that AGI talk tends to skip
Continual learning. A deployed model is frozen. It can hold context in a long window, but it doesn’t consolidate experience into skill the way a person does after a month on a job. Every session starts from the same baseline.
Reliability under novelty. Models handle distribution shift gracefully right up until they don’t, and the failure mode is usually confident invention rather than a clean “I don’t know.” Calibration is still a research problem, not a shipped feature.
Grounding in the physical world. Language models learn from a corpus that dwarfs anything a child encounters, yet children pick up physical intuition from a few thousand hours of stumbling around a room. Sample efficiency of that kind is absent.
Which raises the awkward question of what we’re measuring against. OpenAI’s working definition of AGI, a highly autonomous system that outperforms humans at most economically valuable work, is deliberately economic rather than cognitive, and that framing quietly decides which capabilities count. Reasonable for a company selling labour. Poor for anyone reasoning about machine intelligence in general.
What OpenAI says comes next
Public statements from OpenAI leadership describe a path through agentic systems toward automated AI research: models that propose, run and evaluate their own experiments. The published OpenAI AGI roadmap rewards close reading, partly for the technical claims and partly for which milestones get dates and which stay vague. Compute remains the loudest variable, with training runs scaling roughly an order of magnitude per generation and infrastructure commitments measured in gigawatts.
Timelines diverge wildly for defensible reasons. The assumptions behind the 2027 and 2100 forecasts differ less about hardware than about whether current architectures keep yielding gains or need a conceptual step nobody has taken yet.
Capability and control run on different clocks
Every capability jump raises the stakes on the safety side. OpenAI’s approach leans on scalable oversight, deliberative alignment and chain-of-thought monitoring, backed by a preparedness framework that triggers extra safeguards when a model crosses defined capability thresholds. Whether those thresholds sit at the right levels is a live argument inside and outside the company. The research race in OpenAI’s alignment work and the human review layer that slows it down is the clearest place to watch how that tradeoff is being managed rather than described.
For anyone outside the lab, the practical skill is calibration. When a new capability claim lands, ask three things: what was the eval, who graded it, and how many sequential steps did the task require? Those questions cut through most of the noise, and they’ll keep working long after this model generation is a footnote.

