A year of Gemini progress + what comes next

The announcement has not yet gone live when Logan Kilpatrick introduces a new Gemini 2.5 Pro update onstage. The immediate question is whether this release closes the gaps developers found in earlier versions well enough to become the final update. He calls it “hopefully the final update,” leaving room for further changes if feedback exposes problems.

Kilpatrick calls the update state of the art on Aider and HLE, without supplying scores or evaluation conditions. He connects the improvements to feedback on previous versions and describes 2.5 Pro as a turning point both inside DeepMind and in developers’ perception of Gemini. The complete launch slide also highlights LMArena and WebDev Arena gains, thinking budgets, a stable model candidate, and the Pareto frontier. The release is a candidate for stability, but the invitation to report failures remains open.

For trying the release, Kilpatrick points developers to ai.dev and users to the Gemini app. Those are the access routes he presents during the announcement; the rest of the talk explains how Google reached this point and where the model, app, and developer platform are heading.

The retrospective starts with the density of work packed into the preceding year. DeepMind’s research spans science, robotics, and general models, and Kilpatrick identifies that breadth as an advantage: specialized research can feed back into the mainline Gemini model rather than remaining isolated in separate projects.

He cites AlphaProof and AlphaGeometry as examples of custom-model work improving the mainline models in their respective domains. The important distinction is between developing a specialist system and transferring useful results into a general model. Kilpatrick describes that transfer as happening, but does not specify a training mechanism or quantify either system’s contribution to 2.5 Pro. He leaves a deeper discussion to Jack’s upcoming presentation.

Adoption has expanded alongside the research. Citing Sundar Pichai’s Google I/O keynote, Kilpatrick reports approximately 50× growth in monthly tokens processed across Google products and APIs over a year. The keynote’s totals are 9.7 trillion a year earlier and over 480 trillion at the time of its May 2025 report. This measures aggregate token volume, not inference speed or compute growth. Kilpatrick connects the increase to demand for Gemini within Google and from external developers.

Part of the explanation is organizational. Google previously had separate teams pursuing different strands of AI research. The April 2023 consolidation of Google’s Brain team and DeepMind created a shared organization with a remit extending beyond foundational research: build models and deliver them to the rest of Google and to external users.

Product teams subsequently moved closer to the research organization. Kilpatrick describes a further step earlier in the year of the talk; the Gemini app team’s own move had been announced in October 2024. The resulting structure puts research, model creation, and product delivery inside DeepMind, with the Gemini app serving consumers and the Gemini API serving developers. The slide pairs an organizational update with Kilpatrick’s post about moving AI Studio and the Gemini Developer API to DeepMind.

For Kilpatrick, that proximity makes it easier to collaborate with researchers and bring new capabilities to developers. His formula is direct: bring the best people together, find infrastructure advantages, and ship. The organizational change matters because it connects research progress to product delivery.

Veo provides an example of demand for the resulting products. Kilpatrick jokes that its reception is “burning all the TPUs down,” describing strong interest rather than a measured infrastructure failure. At this point in the talk, Veo is already available in the Gemini app; developer API access appears later in the roadmap.

The Gemini app’s intended destination is a universal assistant. Kilpatrick explains this through a familiar comparison: historically, a Google account connected someone to many Google products, but primarily by signing them into each one. It did not itself provide a rich, stateful experience spanning those products. Gemini is intended to become that connecting thread.

The next step is proactivity. In Kilpatrick’s description, most AI products still make the user initiate the interaction and do much of the work. A proactive assistant would take more initiative, changing not only the interface but also who starts a useful interaction. This remains a product direction rather than a demonstrated workflow in the talk.

For feedback on the app itself, Kilpatrick directs people to Josh and the Gemini app team rather than to him. That division becomes useful as the roadmap separates the consumer assistant from the underlying models and developer tools.

Gemini’s original ambition was a single multimodal model handling audio, images, and video. Native audio and text-to-speech capabilities announced at that year’s I/O move it toward that goal: users can speak with the model, and conversational audio powers the Astra experience and Gemini Live. Kilpatrick describes the destination as an omnimodal model, bringing these capabilities together.

Kilpatrick characterizes Veo as state of the art across several areas, without naming evaluations or scores, and hopes to bring its video capabilities into mainline Gemini. He then points to early diffusion experiments as a route to much higher token generation rates. Diffusion is still a research exploration here, not a mainline capability, and he gives no numerical throughput result.

The next change concerns what a developer receives from a model. Historically, Kilpatrick thought of the interface as tokens in and tokens out, with surrounding software supplying the scaffolding needed to build a larger system. The direction he calls “agentic by default” would move more of that work into the model itself.

Reasoning may absorb some of the work now performed by external agent scaffolding. As models become more systematic, the reasoning step becomes a possible place for that behavior to live. Kilpatrick poses the amount of work that will move as an open question, with consequences for people building products around today’s models. He points again to Jack’s forthcoming discussion of scaling reasoning; he does not announce the replacement of developer orchestration.

The roadmap also includes more small models and more large models. Beyond size, Kilpatrick identifies a harder architectural frontier: infinite context. He argues that the current attention-based model paradigm cannot simply scale indefinitely, so expanding the amount of context models can use will require new innovations. Infinite context is a research ambition, not a promised context-window size. Tulsee is identified as the person leading Gemini model product work and receiving feedback about model capabilities.

The developer roadmap starts with embeddings. They remain important even as attention shifts toward agents because they support retrieval-augmented generation: retrieving relevant material to supply as context for generation. Kilpatrick calls the Gemini embeddings model state of the art, without specifying an evaluation, and expects a broader developer rollout in the next couple of weeks relative to the talk.

The proposed Deep Research API addresses a different layer. Rather than exposing only a model primitive, it would make research workflows—already popular in consumer products—available for developers to build into their own applications. Kilpatrick describes a bespoke API under development, not an endpoint or request format developers can use in this recording.

Veo 3 and Imagen 4 are also planned for API access. Kilpatrick hopes they will arrive very soon, tying availability to the work of scaling the developer platform. These rollout statements describe the roadmap at the time of the recording, not present-day API availability.

The substantive roadmap ends with a positioning decision: AI Studio is being built as a developer platform. Kilpatrick expects it to move away from its consumer-like feel toward an experience designed more explicitly for developers. The accompanying roadmap slide places that focus alongside embeddings, media APIs, agents, a Computer Use API, evals tooling, a Deep Research API, and Gemini Code / CLI.

The intended developer experience includes agents built in. Kilpatrick hopes to integrate Jules and other coding agents natively into AI Studio, making them part of the environment where developers work with Gemini. That is the final product direction: not just a place to try a model, but a platform where model access and assistance with development come together. Native Jules integration remains an aspiration in this presentation.