Google AI Updates: October 31, 2026

1. Project Suncatcher Prototype Returns the First AI Response From a Data Center Class TPU in Orbit

Google reported a “Hello World” milestone for Project Suncatcher, its research moonshot to explore powering AI compute in space. The prototype satellite launched on October 1, and over the following weeks the team powered up the host and TPU trays and ran early machine learning tests. During a ground station pass, engineers uploaded a prompt to a Gemma 4 open model running on the onboard TPU asking for a haiku about being in space, and the response was downlinked as the satellite passed over Svalbard, Norway. Follow-up prompts included limericks and telemetry requests reporting the status of payload systems. Google says the next steps are testing TPU performance under thermal stress and radiation exposure in orbit, with twin interconnected satellites planned for 2027. The post gives no performance, power, or compute figures and names no partners. Source

2. Gemini API Computer Use Reaches General Availability

Google moved computer use to general availability in the Gemini API on October 30, across both the Interactions API and the GenerateContent API, according to the API changelog. The capability lets a model read screenshots and return UI actions such as clicks, typing, scrolling, drag and drop, and navigation, while the developer runs the client-side browser, Android, or desktop environment that executes them; coordinates are normalized to a 0 to 999 range and must be scaled to the viewport. Supported models listed in the docs are Gemini 4 Argon, Gemini 3.8 Flash (the recommended option), Gemini 3.5 Flash-Lite, and Gemini 3 Flash Preview. Built-in safety policies cover financial transactions, sensitive data modification, communication tools, account creation, and legal agreements, each action can carry an allowed, require-confirmation, or blocked decision, and opt-in prompt injection detection requires Gemini 3.5 Flash-Lite or later. The same changelog notes that the stateless, dependency-free gemini-api command line client entered beta on October 27 for scripts, CI pipelines, and coding agents. Source

3. Google Cloud’s October AI Infrastructure Roundup Covers Agent Compute Engine, llm-d 1.0, TorchTPU, and Vera Rubin A5X

Google Cloud published its October roundup of AI infrastructure and orchestration updates. Headline items include Agent Compute Engine, a compute service that runs agents in hardware-isolated microVM sandboxes with sub-second activation and copy-on-write memory sharing, paired with an Agent Gateway and Agent Registry for governance; llm-d 1.0, the open source Kubernetes-native distributed inference stack, with cache-aware routing, tiered KV cache offloading, and priority-based flow control across mixed TPU and GPU fleets; and TorchTPU, a native PyTorch stack for TPUs developed with Meta that supports Hugging Face, vLLM, and TorchTitan. On hardware, Google listed A5X instances built on NVIDIA Vera Rubin NVL72 with the Virgo scale-out fabric and Managed Lustre storage, plus NVIDIA HGX B300 support for air-gapped Google Distributed Cloud. GKE gained a CPU startup boost preview that uses in-place pod resize to cut cold starts by up to 2x, and the roundup also covers ML Drift, a cross-platform on-device GPU engine for LiteRT, day-0 deployment recipes for the Apache 2.0 Reflection Beam model, and customer figures such as ZenteiQ’s 4.07x aggregate inference throughput gain with GKE Inference Gateway and llm-d. Source

4. Gemini in Google Slides Can Now Edit an Entire Presentation From One Prompt

Google began rolling out the ability for Gemini in Google Slides to edit multiple slides or a whole deck from a single prompt in the side panel. Gemini decides which slides need changes, edits the native slides directly, and posts a summary of the changes; the edits can be reverted with undo or version history. Example prompts include matching the style of a reference file, updating a deck with details from a Drive file, summarizing slides 5 through 7 into one slide, and making all titles the same font size and color. The extended rollout started on October 29 for Rapid and Scheduled Release domains, in English only, for Business Standard and Plus, Enterprise Standard and Plus, Google AI Pro for Education, and the Google AI Pro and Ultra consumer plans. Workspace customers get promotional access to higher limits for multi-slide edits through at least December 1, 2026, after which per-user limits apply. Source