Google launches Gemini 4 Hydrogen, replacing 3.8 Flash across Search and API with 2.5x faster completion
·2026.11.13 09:16
Key point
The new model completes tasks 2.5 times faster than its predecessor and scores 57.0% on Terminal-Bench 4.1, while leading competitors in multimodal and reasoning benchmarks.
Details
- Google released Gemini 4 Hydrogen, a speed-optimized variant of Carbon that replaces Gemini 3.8 Flash across Search, the Gemini app, API, and enterprise services.
- Hydrogen completes tasks 2.5 times faster than 3.8 Flash at the default high reasoning setting and achieves a 57.0% success rate on Terminal-Bench 4.1, up from 18.2%.
- On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at high and 54 at xhigh, outperforming Gemini 4 Argon's score of 50.
- In Google's model card comparisons, Hydrogen leads Claude Sonnet 5.5 and GPT-6.1 Terra on MKA (53.4%), HLE-Diamond (58.8%), and SWE-sweep (13.5%), but trails on Terminal-Bench 4.1 (57.0% vs 63.9%) and OSWorld-2.1 (77.6% vs 83.9%).
Search and Developer Impact
- Search now uses Hydrogen for improved query fan-out, allowing broader source discovery and better handling of complex or typo-ridden queries.
- API pricing remains unchanged at $1.50 per million input tokens and $7.50 per million output tokens, with a 1 million token context window.
- Google Antigravity introduced Orchestrator, which routes tasks between models to reduce quota consumption by 24% and time by 19% in internal tests.
Safety and Limitations
- Unauthorized agent actions fell by up to 81% and resistance to indirect prompt injection improved by more than 2.3 times compared to 3.8 Flash.
- Unnecessary refusals of legitimate requests increased by 5.57 percentage points, particularly when prompts include images.
- The model has a knowledge cutoff of March 2026 and requires external tools for more recent information.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.