Google DeepMind and AGI: what is it trying to build?

This page is part of our series on who is building superintelligence. It concentrates on what Google DeepMind says about AGI and beyond, and on the evidence behind it.

The stated objective

Google DeepMind’s About page, checked on 3 October 2026, gives its mission as “to build AI responsibly to benefit humanity”, and says that AI “and ultimately artificial general intelligence (AGI)” has the potential to drive one of the greatest transformations in history. AGI is therefore named as a destination, though the mission itself is phrased more broadly.

Its working definition, published in April 2025 in a post by senior researchers including co-founder Shane Legg, is “AI that’s at least as capable as humans at most cognitive tasks”. That is a cognitive definition, where OpenAI’s is economic (see our OpenAI profile); the difference is one reason the two companies’ statements about AGI are not directly comparable. Our AGI explainer sets out the main definitions.

The company’s leaders now speak of AGI as near. In an August 2026 message announcing a reorganisation (below), Demis Hassabis wrote that he felt AGI “is close at hand”. In January 2026, according to Fortune, he put the chance of AGI within the decade at about 50%, adding that one or two more breakthroughs were probably still needed. These are personal forecasts, not findings; others are collected on our timelines page.

How DeepMind proposes to measure AGI

Google DeepMind’s researchers have published two frameworks for deciding what would count as AGI.

• Levels of AGI (first published November 2023, revised September 2025) rates systems on performance, from “Emerging” to “Superhuman”, and on generality. In its latest version, the general-purpose systems it names (ChatGPT, Bard, Llama 2 and Gemini) sit at the lowest level, “Emerging”, and the higher general levels are listed as “not yet achieved”. The paper does not assess the systems released since.
• A cognitive framework (March 2026) proposes measuring progress against ten human cognitive abilities: perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving and social cognition.

Neither framework says AGI has arrived. Both reflect a view, set out in more detail on our page on recognising AGI, that AGI is better treated as a profile of abilities than as a single line to cross.

From AGI to ASI

In June 2026 fourteen researchers, all at or formerly at Google DeepMind and including Legg and Marcus Hutter, published a paper titled “From AGI to ASI”. It is an unusually detailed public treatment, by researchers at a frontier developer, of what might come after AGI.

The paper defines an AGI as “a system that is roughly as intelligent as a single human”, and sets a demanding bar for artificial superintelligence: a system that exceeds the performance of “large human-expert collectives on virtually all tasks and domains”. That is close to the research sense of superintelligence this site uses, which compares a system with the best humans (explained here); the paper raises the bar further by comparing it with whole groups of experts working together.

It identifies four possible routes from AGI to ASI:

1. scaling up AGI systems;
2. new AI paradigms;
3. recursive improvement, AI improving AI (explained here);
4. superintelligence emerging from large collectives of cooperating AI agents.

On speed, the authors write that recursive improvement “could be super-exponential” and that, if there are no major frictions and AI can improve itself autonomously, the transition from AGI to ASI “may indeed be rapid”, while presenting these as possibilities rather than predictions. The questions they raise are the ones covered on our pages on the intelligence explosion and AI takeoff.

Two cautions. The paper is a research contribution by DeepMind staff, not a statement of company policy; this site found no indication that Google has adopted it as an official position. And it is a map of possible pathways, not evidence that any of them is being travelled. It says comparatively little about safety, beyond noting that a cooperative superintelligence would need careful training and evaluation.

What it has built

Google DeepMind’s best-known demonstrations are in science and mathematics:

• AlphaFold (2020–21) predicts protein structures with accuracy competitive with laboratory methods in most cases; Hassabis and John Jumper shared half of the 2024 Nobel Prize in Chemistry for it (David Baker received the other half).
• Olympiad mathematics: in July 2025 an advanced Gemini model solved five of six International Mathematical Olympiad problems, a gold-medal score officially certified by the Olympiad.
• AlphaEvolve (May 2025), a coding agent that discovered improved algorithms, including one that sped up part of Gemini’s own training.

Its general-purpose models are the Gemini family. On 30 September 2026 it released Gemini 4 Argon, initially only to “trusted cyber defenders” through a programme it calls Fairwind, with wider release, starting with paying API customers and Google AI Ultra subscribers, promised “as soon as possible”. Independent results for earlier models are mixed in the way typical of frontier AI: in September 2026 the ARC Prize Foundation recorded Gemini 3.8 Flash scoring 89.2% on its second benchmark and 10.4% on the harder third using ARC’s standard test harness (35% with a provider-specific harness). Our capabilities page separates what AI has demonstrated from what is extrapolated.

Safety, and the AI-research threshold

Google DeepMind’s main safety policy is its Frontier Safety Framework, now at version 3.1 (April 2026). It sets “critical capability levels” at which extra safeguards apply, for chemical, biological, radiological and nuclear (CBRN) weapons, cyber attack, harmful manipulation, and a combined domain of machine-learning research and development and misalignment. The last is the most relevant here. One threshold is AI that could “fully automate the work of any team of researchers at Google focused on improving AI capabilities”; another is AI substantially accelerating progress beyond historical rates. The framework also tracks whether a model could undermine human control.

Google DeepMind reports against these thresholds. Its August 2026 report on Gemini 3.7 Flash, for example, said the model reached the “alert thresholds” (early-warning levels) for CBRN and cyber capabilities but not the critical levels, and stayed below the critical levels for AI research, manipulation and misalignment. These are the company’s own assessments. It had not published such a report for Gemini 4 Argon as of 3 October 2026.

Its April 2025 paper “An Approach to Technical AGI Safety and Security” sorts risks into four areas, misuse, misalignment, mistakes and structural risks, and concentrates on the first two. Its alignment team’s published priorities in 2026 include monitoring models’ written reasoning, debate-based oversight and interpretability. Our risks page and control page explain these approaches.

One incident has been reported: in September 2026 Google confirmed, according to the Wall Street Journal, that a Gemini model had broken into the systems of three real companies during a cyber-security test run by an outside firm in May. Google’s vice-president of security engineering, Heather Adkins, said that “the model acted appropriately”, stopping once it realised it had reached real rather than simulated systems. The test firm, Irregular, attributed the incident to a fictional test domain that matched a real one. This site has not seen a primary account.

Organisation

Google DeepMind was formed in April 2023 by merging DeepMind, founded in London in 2010, with Google Brain. It is part of Alphabet, and its headquarters are in King’s Cross, London. On 5 August 2026 Google announced that Hassabis would become chair of Google DeepMind and chief scientist of Alphabet, and that Koray Kavukcuoglu would run Google DeepMind day to day as senior vice-president, while also serving as Google’s chief AI architect. Hassabis also continues to lead Isomorphic Labs. As of 3 October 2026, Google DeepMind’s own About page still described Hassabis as “Co-founder and CEO”. Legg remains co-founder and chief AGI scientist. In September 2026 Legg, Hassabis and James Manyika launched the DeepMind Institute, publishing essays it says should not be read as Google’s official view. Our UK page explains what DeepMind’s London base does and does not mean for Britain.

Google also signed the voluntary White House Accord on Super Intelligence on 29 September 2026, through its chief executive Sundar Pichai.

Claimed, demonstrated, unknown

• Stated aim: AGI, as the culmination of building beneficial AI; its leaders say AGI is close.
• Demonstrated: Nobel-recognised scientific AI (AlphaFold), certified olympiad mathematics, and frontier general-purpose models with independently measured strengths and weaknesses.
• Theoretical: the routes from AGI to ASI in its researchers’ 2026 paper.
• Unknown: how its newest systems score against its own AI-research thresholds, and whether the frameworks it has proposed for recognising AGI will be the ones it uses when the question arises.