HiRO-ACEv2 opens 3 km climate runs, Earth AI and Deep Review

No big frontier launch landed in my window today. What did land is a quieter set of releases that I'd actually use, so I'm leading with the one you can download.

HiRO-ACEv2 puts 3 km climate runs in reach

Allen AI released HiRO-ACEv2, an emulator that combines ACE, which runs at 100 km resolution, with HiROv2, a downscaling model that adds detail down to 3 km. Their number: global 3 km output at about 1/300th the cost of X-SHiELD. They say it matches the target simulation's statistics within 3% for extreme events, and that it reproduces tropical cyclone energy and realistic hurricane structure. Over 10-year runs, the hurricane energy lines up with the reference data.

The part I care about is the release. Weights are on Hugging Face, there are tutorial notebooks, and a web demo lets you look at output without running anything. Cheap, open, high-resolution climate simulation is the kind of thing that changes who gets to do the research. These are the lab's own claims, and I haven't run it.

Agent plumbing keeps moving

OpenAI's Python SDK hit 3.31.0 with browser and computer toolsets. The release note labels it part 1 of 4, so expect more to follow. It's a small diff today and a large one if the other three parts arrive.

NVIDIA published a post on agent reliability and tool-calling efficiency for chip design using NeMo Gym. The framing is the honest one: agents are productive in complex design work, but they generate plausible, wrong artifacts. Anyone who has watched an agent confidently fill a spreadsheet with bad numbers knows the problem.

ElevenLabs wrote up how it handles 22,000 support tickets a month with ElevenAgents. The agent looks up account information, takes actions and escalates, built on procedures, simulation tests and tools. I like that simulation tests are in the list. That's the part most teams skip.

Google Research, two science releases

Earth AI now has a geospatial reasoning agent, reachable through Ask Google Earth and an Agent-to-Agent API, with two new benchmarks, GeoKTools and GeoExpertQueries. Google cites WHO AFRO mapping 48 exposed settlements and 45,500 at-risk people during the 2026 DRC Ebola outbreak in minutes rather than weeks. Early access is by interest form.

Deep Review is the one you can try today. It's an agentic system for checking technical documents: segmentation, adaptive budgeting, parallel verification, synthesis. Google piloted it on more than 16,000 manuscripts across STOC, ICML, NeurIPS and ICLR. You get two free reviews a week, then you bring a Gemini API key.

What I'd watch next

Whether parts 2 to 4 of the OpenAI toolsets show up this week. Whether anyone outside AI2 publishes numbers on HiRO-ACEv2. And whether Deep Review holds up on documents that aren't academic papers, since that's where I'd want to use it.