Google ships Gemini 4 Hydrogen: beats Argon in benchmarks at ~32% of Sonnet 5.5's cost
APPSO · wechat · 2026-11-13
Google released Gemini 4 Hydrogen, a fast, low-cost model aimed at agent execution: rapid coding, tool calls, and multi-step tasks. On Artificial Analysis it scores 51 at default High, beating Gemini 4 Argon's 50, and 54 on the new xhigh tier—just 1 point below Claude Sonnet 5.5 and GPT-6.1 Sol at max effort. API pricing is $1.50/$7.50 per million input/output tokens, uniform across the 1M-token context; xhigh costs 32% of Sonnet 5.5's top-tier per-task cost. Task speed at High is 2.5x Gemini 3.8 Flash, and hallucination rate is just 8% vs 43%/52% for the two flagships.
Community demos show one-prompt 3D harbors, Three.js cities, and playable games; Google's Geminisland generates voxel buildings in real time. Antigravity gained Orchestrator, which splits work between Carbon (planning) and Hydrogen (execution) with parallel agents, cutting quota use 24% and speeding tasks 19% on Terminal-Bench 4.0. Hydrogen rolls out globally to Gemini App, Google Search, and AI Studio, including free users. The piece argues Hydrogen may be the most underrated Gemini 4 model—it decides how many people can actually afford these AI capabilities.
More from coding & agent
- Sketchy Update: Claude Now Sketches Molecules and Circuits, Open Source — Nkgsporty · 2026-11-13
- Dev postmortem on Opus-reviewed fixes: ~1/3 accepted as-is, blind trust would've been horrible — Vjeux · 2026-11-13
- Mediabunny bughunt done: nearly 400 fixes, prize for anyone finding a remaining bug — Vjeux · 2026-11-13
- Descope's Envoy gives AI agents their own sign-in identities, compatible with Poppy — AiBreakfast · 2026-11-13
- Beaker makes production agents 37% faster and 35% cheaper, claims the model isn't the bottleneck — VictorSanh · 2026-11-13
- CMU trains LLMs to write memories with RL for continual learning across task streams — aviral_kumar2 · 2026-11-13