Tech • AI • Robotics • Game
Google Just Dropped HYDROGEN: The New King of Everyday AI
9.4/10TL;DR
Google has launched Gemini 4 Hydrogen, a lower-cost, faster AI model aimed at coding and agent workflows, while Microsoft is teasing new models ahead of Ignite and OpenAI has begun testing agent-driven shopping through ChatGPT Wallet.
KEY POINTS
Google targets everyday AI work
Gemini 4 Hydrogen launched on November 12 as a lightweight model built for speed, coding, and autonomous multi-step tasks. It sits below Google’s flagship Carbon model and is positioned against Anthropic Claude Sonnet 5.5 and OpenAI GPT 6.1 Terra, the models commonly used for day-to-day professional workloads. Google is rolling it out across Search, the free Gemini app, AI Studio, and the Gemini API.
Benchmark results favor Hydrogen on several practical tests
On Swiss Sweep, a code-fixing benchmark using real software issues, Hydrogen scored about 13.5%, compared with 5.8% for Terra and 3.1% for Sonnet. On Agents Last Exam, which measures completion of end-to-end professional workflows, Hydrogen reached nearly 72%, ahead of Terra at just under 70% and Sonnet at about 60%. It also led on Automation Bench at roughly 76%, versus 72% for Sonnet and 60% for Terra.
Strengths come with a few notable weaknesses
On MKA, a multimodal test involving visual files, Hydrogen posted about 53%, narrowly ahead of Sonnet at 52% and clearly above Terra near 44%. But Google’s own comparisons show Hydrogen trailing Sonnet on OSWorld 2.1, where models control a computer interface, and falling behind both rivals on Terminal Bench Science. The picture is of a model that is highly competitive in office automation and coding, but not dominant across every category.
Google’s main pitch is speed and cost
Google says Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash on its default high setting. It also offers four reasoning modes, from low to X high, allowing users to trade speed for quality. Third-party evaluator Artificial Analysis gave Hydrogen an Intelligence Index of 51 on its default setting and 54 at X high, compared with 55 for Sonnet 5.5 at max and 51 for Terra at max.
Price could be the biggest competitive lever
Cost figures cited by Artificial Analysis show Sonnet at its highest setting costing about $10.02 per task, versus $3.16 for Hydrogen at its highest setting. That places Hydrogen at roughly a third of Sonnet’s cost while giving up only a small amount on top-end benchmark performance. The result is a model that may appeal more to businesses running large volumes of daily coding, support, and document work than to buyers focused only on leaderboard wins.
Demos focus on 3D worlds, animation, and media production
Google showcased Hydrogen generating a walkable 3D island game from a single prompt, with roads, buildings, and large animated creatures. Other examples included a 3JS canyon flight game, a real-time physics and light simulation, and an automatically assembled video using public footage, editing, and voiceover. In one head-to-head example, Hydrogen completed an animated SVG pelican riding through five seasons in about five minutes, compared with roughly 25 minutes for Fable 5.2.
Google is pairing Hydrogen with Carbon for software agents
Through a new Orchestrator feature in Google’s coding tool Anti-Gravity, Carbon handles planning and codebase analysis while Hydrogen performs the execution work such as editing files, updating dependencies, and running tests. On Terminal Bench 4.1, Orchestrator achieved nearly the same completion level as Carbon alone while using 24% less quota and finishing 19% faster on average. Google’s framing is a senior-junior pairing: Carbon plans, Hydrogen executes.
Microsoft is preparing its next AI reveal
Microsoft Ignite, scheduled for November 17-20, is expected to feature new MAI models, demos, and sessions, though the company has not provided names or specifications. Earlier this year, Microsoft introduced MAI Thinking One, a mixture-of-experts model with about 35 billion active parameters per answer and a 256,000-token context window. Microsoft said the model scored 97% on AIM 2025 and 53% on Swebench Pro, while its MAI Code One Flash reached 51% on Swebench Pro with only 5 billion parameters.
OpenAI starts testing AI-powered checkout
On November 13, OpenAI introduced ChatGPT Wallet for a limited set of U.S. ChatGPT Pro subscribers aged 18 and older using a U.S.-issued Visa, Mastercard, or American Express. Users can store a card once, then let Dots, OpenAI’s shopping agents, select items and proceed to checkout after biometric approval. Payments rely on tokenized card details, keeping the actual card number hidden from both the agent and the merchant.
Retail resistance could shape the future of shopping agents
OpenAI plans to add spending limits and category-based rules so purchases can eventually be preapproved and delegated. But scaling that model depends on merchant cooperation, and some retailers are pushing back. Amazon blocks outside AI shopping agents while steering customers toward Alexa, highlighting a likely battle over who controls agent-based commerce at the point of sale.
CONCLUSION
The latest AI competition is shifting from headline benchmark wins to a three-way fight over speed, operating cost, and control of real-world workflows. Google, Microsoft, and OpenAI are now competing not just on model quality, but on who can own the software agent layer that increasingly acts on a user’s behalf.
Ask a question