AI Today: Gemini 4 Hydrogen and OpenAI Agents
Authorship and approval
Published and updated
State of knowledge as of 2026-11-17. Source material gathered on 2026-11-17.
Today's topics cover Google's Gemini 4 Hydrogen with a million-token context window, OpenAI medical agents, and a cluster of four DGX Spark systems. Google is also expanding Ask Maps with trip planning, while OpenAI makes its CSAM classifier available to organizations. The materials describe test results and feature reach, but in a few cases do not specify availability in Poland or costs.
Highlights Today (TL;DR)
- Gemini 4 Hydrogen: the model offers a context window spanning one million tokens.
- OpenAI medical agents: in tests, 99% of 325 tasks were rated as safe.
- NVIDIA DGX Spark: four systems form a cluster with 512 GB of memory and 20 TB of NVMe storage.
- Ask Maps: Google adds trip planning and availability checks for saved locations.
- OpenAI CSAM classifier: in tests, it could reduce the number of images to review by 48% with a similar number of detected materials.
Google Releases Gemini 4 Hydrogen with a Million Tokens of Context
Google has released Gemini 4 Hydrogen for fast coding and AI agent workflows, where programs execute sequential tasks. The model features a context window of one million tokens — units of text that the model considers all at once.
- API pricing: 1.50 dollars per million input tokens and 7.50 dollars per million output tokens, regardless of context length.
- Artificial Analysis Intelligence Index: 51 points in default mode compared to 50 for Argon; the xhigh mode reaches 54.
- An orchestrator that selects models during operation achieved results close to Carbon on Terminal-Bench 4.0, consuming an average of 24% less limit and running 19% faster.
- Compared to Gemini 3.8 Flash Hydrogen, it completes tasks 2.5 times faster; the score on Terminal-Bench 4.1 rose from 18.2% to 57%, and the number of hallucinated claims dropped more than fivefold.
- The model is available in Google AI Studio, Gemini API, Gemini for Business, the Gemini app, and Google Search, as well as for paid AI Pro and Ultra subscribers in Antigravity.
In the lab: On October 14, we wrote about Gemini 4 Argon as a model for the hardest problems; Hydrogen is designed to handle fast task execution in this family.
For AI builders: Development teams can test Hydrogen on their own tasks in Google Antigravity. The source does not state EU availability, Polish language quality, or usage limits beyond token pricing.
OpenAI Releases Medical Agents, 99% of Tasks Rated Safe
OpenAI has released computer-use AI agents for healthcare; in tests, 99% of 325 tasks were rated as safe. The agents navigate between systems, click, and enter data, preparing documents for review by clinicians and administrative staff.
- Doctors rated the usefulness of GPT-6 Astra's work as "very" or "extremely" high in 93% of 646 ratings.
- In the administrative tasks benchmark, GPT-6 Astra scored 94.3% points, and GPT-6 Luna 91.8%; Luna cost 8.5 cents per task, 98% less.
- In a synthetic data demonstration, the agent checked patient documentation and prepared an MRI order and prior authorization request. A doctor and specialist still needed to approve them.
- Access is available via the OpenAI API, ChatGPT for Healthcare, and ChatGPT for Clinicians. Organizations can set agent approvals and access as well as maintain audit logs.
- CVS Health used the agent to catch up on a backlog of document submissions that had accumulated over months during the weekend.
In the lab: The lab database features information on the deployment of ChatGPT at AdventHealth to streamline work; the current announcement adds agent test results and specific administrative processes.
For AI builders: Healthcare facilities and medical software vendors should check how to configure action approvals and auditing. The announcement does not state deployment costs or availability in Poland or the EU.
NVIDIA Introduces Four-Node DGX Spark with 512 GB of Memory
NVIDIA has introduced the HPC Developer Kit featuring four DGX Spark systems connected in a cluster. Researchers and developers can run multi-node scientific computing and AI models right on their desks instead of waiting for a shared GPU cluster.
- Four systems are connected in a ring via a ConnectX-7 network; no additional network switch is required.
- 512 GB of memory and 20 TB of NVMe storage make up the cluster's combined resources.
- Models run locally with NVFP4 — a quantization format for lower-precision model storage — can reach up to 700 billion parameters. The kit includes four DGX Spark systems; the source does not specify memory requirements for other configurations.
- NVIDIA HPC SDK is slated to receive installation instructions for the cluster soon. The kit is already offered through OEM partners and the NVIDIA Marketplace; the source does not state the price.
In the lab: Previously, we wrote about DGX Spark as hardware for local AI. Now, NVIDIA is showcasing four systems connected for collaborative work.
For AI builders: Research teams and developers considering a purchase should check the price and availability in Poland — the material does not provide them. HPC SDK installation instructions have not yet been released.
Google Expands Ask Maps with Travel Plans and Time Slot Search
Google is adding route planning, free time slot searching, and saved location availability checks to Ask Maps. The service is rolling out to 30 new European countries, though some features launch first in the US.
- Plans will include day-by-day itineraries and numbered pins; they can be shared or exported to Google Docs on a computer.
- Driving directions will display features such as traffic, weather, toll fees, and parking.
- Agentic capability — AI performing tasks rather than just answering — will search for service time slots and can call local businesses. It launches first in the US.
- Saved place search will check their current availability; trip planning and this feature are available wherever Ask Maps works.
- Updates are rolling out to Android and iOS. Ask Maps is expanding to 30 European countries and now covers over 200 countries and territories.
In the lab: We wrote about Gemini in Apple CarPlay, which enables voice conversations in the car; Ask Maps brings part of trip planning and errand running into Maps.
For AI builders: Travel and local service companies should check which features work in their market — time slot searches and route directions start in the US. Google does not state prices or the availability of these features in Poland.
OpenAI Releases CSAM Classifier That Can Reduce Flag Count by 48%
OpenAI is providing verified organizations with free API access to an image classifier that detects potential child sexual abuse material (CSAM), including AI-generated content. Initial tests suggest fewer images to review with a similar detection rate for abuse.
- In tests, the tool could flag 48% fewer images than the previous system while detecting a similar amount of material.
- Flagged images were about twice as frequently material containing abuse.
- Fal confirmed 72% of high-confidence flags as new CSAM; InVideo banned one out of every four flagged instances.
- In SafetyKit tests, the classifier detected 90% of drawings and animations sexualizing children.
- The classifier evaluates visual indicators of abuse, and its results help reviewers determine which images to check first.
In the lab: The lab database contains an earlier OpenAI material on combating online child sexual exploitation and abuse from August 10; the current announcement adds a specific classifier and a signal indicating AI-generated content.
For AI builders: Platform safety teams and child protection organizations should check access criteria and how to integrate the results into reviewer workflows. The announcement does not state API limits or license terms.
Frequently Asked Questions (FAQ)
How much does the Gemini 4 Hydrogen API cost?
Gemini 4 Hydrogen costs 1.50 dollars per million input tokens and 7.50 dollars per million output tokens, regardless of context length. The model is available in Google AI Studio and the Gemini API, among others.
What does the NVIDIA DGX Spark kit include?
The kit includes four DGX Spark systems connected in a ring via ConnectX-7, without an additional switch. It provides a total of 512 GB of memory and 20 TB of NVMe; models up to 700 billion parameters with NVFP4 can run locally.
How does the OpenAI CSAM classifier work?
The classifier evaluates visual indicators of abuse and helps reviewers prioritize image checks. In tests, it could flag 48% fewer images than the previous system while maintaining a similar detection rate for materials.
Sources
- Introducing Gemini 4 Hydrogen — 2026-11-12
- Advancing computer-use agents for healthcare teams — 2026-11-16
- NVIDIA DGX Spark Developer Kit Turns a Desk Into a Supercomputing Lab — 2026-11-16
- New Ask Maps features to help with holiday travel and prep — 2026-11-16
- Strengthening the fight against online child sexual abuse with new tools and support — 2026-11-16
--- Text prepared automatically by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Version before human editing.
Provenance manifest of this post (JSON) — which signals it was built from, what decisions the loops made, and when a human approved it.
Related posts
AI Today: GPT-6.2 Astra and Gemini
Discover the latest updates on GPT-6.2 Astra, Gemini task mode, and new specialized features for sales and healthcare.
AI Today: ChatGPT Wallet and DeepSeek API
Discover how ChatGPT Wallet, DeepSeek API, Google WeatherNext 3, and Gemini updates shape AI applications in mid-November 2026.
API Architecture and Ecosystem in Agentic Systems
An analysis of API architectures, standardization protocols, and performance optimization in modern agentic AI systems.
AI Today: Qodo, OpenAI Pando, and Gemini 4
Stay updated on the latest developments in developer tools, infrastructure, and new artificial intelligence models from Qodo, OpenAI, and Google.