The DeepMind AGI timeline is not a countdown clock. Ask Google DeepMind when artificial general intelligence will arrive and you get a range, a caveat, and usually a sentence about safety. That is not evasiveness so much as the state of the field. Nobody has a schedule, but the people building AGI keep leaving markers.
Demis Hassabis, DeepMind’s CEO, has spent the past few years refining his public estimate. In 2023 he said AGI could arrive “within a decade.” At Mobile World Congress in 2024 he put it at five to ten years. Shane Legg, co-founder and chief AGI scientist, has used 2028 as a rough 50/50 marker for years. Those two numbers, 2028 and the early 2030s, dominate the DeepMind AGI timeline conversation. They are not the same claim.
Why DeepMind doesn’t publish a single AGI date
Google DeepMind is a research lab inside a company that ships products. A public date would become a product deadline, a safety liability and a recruitment slogan. Instead, DeepMind publishes capability frameworks. Its 2024 “Levels of AGI” paper describes performance, generality and autonomy from Emerging to Virtuoso, plus an ASI level above that. It is a scorecard, not a calendar. For a closer look at how the plan is packaged, this analysis of DeepMind’s AGI plan separates the science from the spin.
The lab’s public language is deliberately loose. “Within a decade” can mean eight years or twelve. “Five to ten years” gives a five-year spread. That flexibility is useful when the underlying work is uncertain. It is less useful if you are trying to plan a product, a policy or a career.
The Hassabis range: five to ten years, with a moving start date
Hassabis has been consistent about the shape of the timeline, not the date. In 2023 he talked about AGI within a decade. In 2024 he narrowed it to five to ten years, while saying it could come sooner. The range has stayed stable even as the starting year moves. If someone says “five to ten years” every year, the implied end date drifts. That is a useful tell.
His estimate also carries a condition. Hassabis has argued that AGI must be built with scientific rigor and that safety and alignment need to be solved before deployment. That makes his timeline a research estimate crossed with a risk assessment. The technical work might land on one schedule. The safety case might land on another.
Shane Legg’s 2028 marker is older than Gemini
Legg’s 2028 prediction predates the current boom. He has repeatedly described it as a rough 50% probability, not a promise. In a 2023 interview he put the chance of AGI by 2028 at about 50% and stressed the uncertainty. The number sticks because it is specific. It also contrasts neatly with Hassabis’s wider range. Legg’s date is about research progress. Hassabis’s range includes deployment, safety and societal readiness.
Neither man is giving a launch date. They are giving probability estimates with different assumptions. That distinction matters when headlines collapse “50% by 2028” into “AGI in 2028.”
What would have to happen by 2028
A 2028 arrival would require several hard problems to fall in the same window:
- Long-horizon reasoning. Systems that can plan over days or weeks, not just answer a prompt or finish a coding task.
- Continual learning. Models that acquire new skills without a full retrain and without forgetting earlier ones.
- Reliable autonomy. Tool use and computer control that works across unfamiliar software, including failure recovery.
- Safety cases at scale. Interpretability, oversight and red-teaming that can keep up with capability gains.
- Compute and energy. Enough infrastructure to train and serve frontier models without choking on cost or power constraints.
DeepMind’s current systems show pieces of this. Gemini models handle multimodal input, long context and agentic workflows. AlphaFold solved protein structure prediction, a narrow but superhuman result. Project Astra explores real-time multimodal assistants. None of those is AGI on its own. The timeline depends on whether these pieces compound or hit a wall. The research base is broader than one model family; this review of Google’s AGI research papers tracks the paper and TPU trail.
The product clock and the safety clock
DeepMind’s AGI timeline is pulled by two clocks. The product clock is visible: Gemini releases, API pricing, enterprise deals and features that ship every few months. The safety clock is slower. DeepMind’s Frontier Safety Framework sets critical capability levels and evaluation triggers. It is a governance tool, not a forecast. But it changes the timeline because a model that crosses a threshold may be delayed for review.
Compare that with OpenAI. OpenAI’s AGI timeline has followed a similar pattern of public dates and internal scorecards, with its own definitions and loopholes. The two labs are not running the same race, but they are reading the same stopwatch.
Definitions decide the date
AGI is not a single benchmark. It is a bundle: generality, autonomy, skill acquisition and economic impact. Change the definition and the timeline moves by years. OpenAI’s 14-word definition is a good example of how much work a short sentence can hide; this breakdown of OpenAI’s AGI definition shows why “highly autonomous systems that outperform humans at most economically valuable work” is harder than it sounds. DeepMind’s level-based framing is more explicit, but it still leaves judgment calls.
Ask whether ChatGPT is AGI and you get the same definitional split; the answer depends on who you ask. A benchmark win is not general intelligence. A demo that handles a novel task is not a system that can acquire any skill. The DeepMind AGI timeline will stay fuzzy as long as the target is fuzzy.
Signals that would actually move the DeepMind AGI timeline
Ignore the countdown graphics. Watch for changes in language and evidence:
- Hassabis drops “five to ten years” for a narrower window.
- Legg stops using 2028 as a rough 50/50 marker.
- DeepMind publishes a safety case for a frontier model, not just a framework.
- Gemini or a successor demonstrates long-horizon autonomy with minimal supervision.
- Evaluation suites stop saturating, and new benchmarks survive more than a few months.
- Energy and data constraints force a shift from scaling to algorithmic efficiency.
Watch the language as much as the demos. “Within a decade” is a soft claim. “By 2028” is a bet. “We have a system that can do X without a human in the loop” is an evidence claim. The DeepMind AGI timeline will tighten when those statements start to line up.
The date is a coordination problem, not just a research problem
Even if the technical work lands in 2028 or 2030, deployment will not be a single switch. DeepMind operates inside Google, with cloud customers, regulators and safety reviewers. A model that meets an internal AGI threshold could be held back by external review, compute limits or liability. That is why the most useful way to read the DeepMind AGI timeline is as overlapping bets: a research bet on algorithms, a compute bet on infrastructure, and a governance bet on safety.
For now, the honest answer is a range. Hassabis says five to ten years. Legg points at 2028. The gap between those numbers is where the real work sits.

