Ask Google for its AGI roadmap and you get something quite different from what OpenAI hands out. There is no tidy five-level chart on a public landing page, no numbered ladder with names like “Level 3: Agents.” Instead, the plan is spread across research papers, safety frameworks, keynote demos, earnings calls and hiring pages. Reading it properly means knowing which of those sources actually constrains what Google builds next, and which are just there to look good in a slide deck.
What follows is an attempt to assemble the real thing: the sequencing Google appears to be following toward AGI, the parts that are concrete and engineering-driven, and the parts that are still a press release wearing a lab coat.
Where the Google AGI roadmap actually lives
Three documents or document types do most of the work. First, DeepMind’s published research, which tells you what problems the organisation thinks are solvable and roughly in what order. Second, the Frontier Safety Framework, which is the closest thing Google has to a formal gate on capability. Third, the product release cadence for Gemini, which functions as a public commitment device: once a capability ships to billions of users, the internal roadmap has to accommodate it.
Anyone wanting the paper-level view of this should look at Google’s AGI research output, the TPUs behind it, and the parts that aren’t marketing. The pattern that emerges is consistent: DeepMind publishes the method, Google ships a scaled version of it six to eighteen months later, and the roadmap’s next step is usually visible in a paper before it is visible in a product.
Gemini is the de facto roadmap, generation by generation
Google doesn’t announce milestones like “we have reached reasoning.” It announces model versions. Those versions are the roadmap, and each one adds a specific competence that the previous one lacked.
What the ladder has added so far
- Gemini 1.0 (Dec 2023): natively multimodal from the start, with three size tiers aimed at data centre, cloud and on-device deployment.
- Gemini 1.5 (Feb 2024): mixture-of-experts architecture and a context window that went from 128,000 tokens to 1 million, then to 2 million for developers.
- Gemini 2.0 (Dec 2024): explicitly built around agents, with native tool use, multimodal output and Live API support.
- 2.5 and the “thinking” line (2025): models that spend compute on internal reasoning before answering, plus hybrid reasoning toggles so the same model can be fast or deliberate.
Each step maps to something on the AGI checklist: longer context is a proxy for memory and document-scale grounding; native multimodality is a proxy for real-world perception; agentic tool use is a proxy for taking action; extended thinking is a proxy for planning. If you want a longer treatment of how those proxies hold up, what Google actually means when it ties Gemini to AGI is worth reading alongside this piece.
Compute is the part Google never downplays
Most labs are vague about compute. Google is not, because it owns the stack. The TPU line has moved from v4 to v5p to Trillium (v6e), with Ironwood following as a chip explicitly designed for the inference-heavy, reasoning-model era rather than training alone. Alphabet guided to roughly $75 billion in capital expenditure for 2025, with a large share going to servers and data centres.
The energy side is equally concrete. Google signed a deal with Kairos Power in late 2024 for small modular reactors, with the first unit targeted for 2030 and a fleet planned through 2035. It has also signed nuclear and renewable agreements to feed specific data centre campuses. Scaling laws are a claim about loss versus compute; the roadmap’s compute commitments are the closest thing to a public bet that the claim still holds.
The safety framework runs on its own schedule
The Frontier Safety Framework, published in May 2024 and revised since, sets out Critical Capability Levels across domains including cyber offence, biological risk, machine learning R&D and deceptive alignment. Models that cross a threshold trigger extra internal review before deployment. That structure matters for the roadmap because it creates an explicit gate between a capability existing in the lab and that capability reaching users.
How those gates actually work in practice, and where they remain thin, is covered in detail in this analysis of DeepMind’s AGI safety frameworks and the hard problems behind them. The short version: the evaluation methodology is genuinely serious, and the deployment decisions are still ultimately made by the same people who want to ship.
Agents are the current front line
If you want to know what phase of the Google AGI roadmap we are in right now, look at what the demos are about. Project Astra, first shown at I/O 2024, is a real-time multimodal assistant that watches through a camera and talks back. Project Mariner drives a browser. Gemini Live handles conversational interaction. Gemini Robotics, announced in March 2025, puts Gemini into physical bodies via a vision-language-action model.
The roadmap logic is legible: perception was solved first, then reasoning, and now the work is on sustained action over time. An agent that can reliably complete a twenty-step task with a 95% per-step success rate still fails most of the time. That arithmetic, not any philosophical objection, is what currently caps agent deployment.
The gaps that the roadmap doesn’t resolve
Continual learning
Models still learn mostly at training time. Google’s context windows are a workaround, not a solution: a 2 million token window lets a model read a lot, but it doesn’t update weights from experience. Whether the next Gemini generation introduces real online learning is the single biggest open question in the whole plan.
Long-horizon reliability
Benchmarks like SWE-bench and ARC-AGI measure short bursts of competence. AGI requires thousands of coherent steps. Google’s own research on test-time compute suggests progress here is a matter of scaling deliberation, which is expensive at inference and therefore a business-model question as much as a technical one.
Evaluation
There is still no accepted test for general capability. Google evaluates against internal suites and third-party benchmarks, but a company grading its own homework on a problem it defined is a structural weakness nobody has solved.
Science as the real target
DeepMind’s stated purpose has always been solving intelligence to advance science. AlphaFold, AlphaProof’s silver-medal performance at the 2024 International Mathematical Olympiad, and AlphaEvolve’s algorithm discovery work are the proof points it points to. The roadmap’s success criterion may end up being scientific output rather than a chatbot. The specifics of that research programme are laid out in this look at DeepMind’s AGI breakthroughs and what’s still missing.
Reading the timelines without getting spun
Demis Hassabis has said AGI is likely “a decade” away, a formulation he has repeated with minor variation since 2023. Shane Legg, DeepMind’s chief AGI scientist, has given 2028 as roughly a 50% point for human-level machine intelligence. Those are not contradictory statements; they describe different confidence levels and different definitions.
The useful signal is not the date. It is the sequence. Watch for these, in this order:
- Reasoning models that hold accuracy across hundreds of sequential steps without human intervention.
- Memory that persists across sessions and updates from experience, not just context.
- Agents that can be handed a vague goal and return a finished result rather than a plan.
- Frontier Safety Framework thresholds actually triggering a delayed release, publicly acknowledged.
- Robotics deployments outside demo videos and into manufacturing or logistics partners.
Google’s advantage in this race is not any single model. It is vertical integration: chips designed in-house, data centres it builds itself, a research lab with a decade of foundational work, and distribution through Search, Android and Workspace that no competitor can match. That combination means the roadmap can absorb a disappointing model generation without losing ground, which is a luxury most labs don’t have.
The honest assessment of the Google AGI roadmap is that the compute and safety scaffolding is more real than most people assume, the product cadence is a genuine commitment mechanism, and the hardest problems, continual learning and long-horizon reliability, remain genuinely unsolved rather than merely undiscussed. Anyone tracking this closely should treat the papers as the leading indicator and the keynotes as the lagging one. For a fuller picture of how the strategy fits together day to day, the science, the spin and what’s actually shipping inside DeepMind’s AGI plan fills in the organisational detail this overview skips.

