Google launches Gemini 4 Hydrogen: fast model for agents at $1.50/M tokens

Leer este artículo en español

Google has released Gemini 4 Hydrogen on November 12, just 48 hours after Gemini 4 Carbon. Hydrogen is the «fast model for everyday execution», adapted from Carbon, and closes the Gemini 4 family: Argon (general balance, October), Carbon (deep reasoning and frontier cybersecurity), Hydrogen (agents and daily code). It affects you if you develop agent flows, migrate from 3.8 Flash, or evaluate inference costs at scale.

Unified 1M-token window and long decode change the economics of long-running agents

Hydrogen shares a 1,048,576-token window across input, thinking, and output. The default output is 65,536 tokens, but with long decode in background reaches 1M output tokens. Previously you had 1M input and 64K output separate; now the same budget covers large artifacts (code, reports, 3D visualizations) in a single conversation. Note: each turn charges the full input tokens.

Configurable thinking with high default: adjust cost and latency

Four levels: low, medium, high (default) and xhigh (new). In 3.8 Flash the default was medium. Raising to xhigh raises the AA Index from 51 to 54, but consumes more tokens and latency. For simple tasks, low or medium reduce cost. Configure it explicitly in the API call.

Benchmarks: wins in its class but the regression in multimodal refusals forces monitoring

Google reports Terminal-Bench 4.1 at 57% versus 18.2% for 3.8 Flash, reduces fabricated claims >5× and is 2.5× faster in setting high. Leads its class in HLE-Diamond (58.8%), Vals Tax Agent Bench (52.3%, #1 global), Vals Legal Research Bench (55.8%, tie #1) and AutomationBench-AA. Agent safety: overreach -81%, under-service -79%, misaligned actions -12×, honesty 96.7%, CoT readability 99.9% without deceptive edits in >200k turns, resistance to indirect injection >2.3× (Gray Swan IPI).

The model card acknowledges a real regression: unjustified refusals rise 5.57 percentage points vs 3.8 Flash, «especially in multimodal prompts». Image-to-text improves 7 pp in safety. If your app uses images, PDFs, or audio intensively, check production logs for false positives.

Orchestrator in Antigravity combines Carbon and Hydrogen to save quota and time

The Orchestrator dynamically alternates Carbon (coordinator/auditor) and Hydrogen (workers/verifiers). In Terminal-Bench 4.0 it achieves success rates close to Carbon while consuming 24% less quota and 19% less time. The native multi-model strategy (/swarm) allows maintaining quality on mixed tasks without paying for Carbon for everything.

What to do now: migration, Fairwind, and scientific program

1. Evaluate 3.8 Flash → Hydrogen migration with your real workloads (latency, cost/1k accepted tasks). Partners report 2–2.8× speedup and -12% to -23% tokens, but Google does not detail test conditions.
2. Use Orchestrator/swarm in Antigravity for mixed workloads: Carbon for architecture and deep debugging, Hydrogen for writing, tests, and verification.
3. Configure thinking level explicitly; do not assume the default high if your tasks are simple.
4. Monitor unjustified refusals in multimodal; the 5.57 pp increase is significant in production.
5. Leverage the unified 1M window + long decode for agents that generate large artifacts, accounting that each turn reloads the input.
6. Enroll in the Fairwind Program if you audit open source code or critical infrastructure; Carbon demonstrates 100% recall and 95% precision in scanning and triage, with >25% relative improvement vs Argon in detection of real logic flaws.
7. Researchers: apply to the scientific program (10k spots, 6 months free Argon+Carbon in Antigravity) for formal verification, structural prediction, enzymatic assays, etc.

Hydrogen’s flat pricing ($1.50/M input, $7.50/M output) doubles current 3.8 Flash, but 3.8 Flash will rise to that level on January 1, 2027. Knowledge cutoff: March 2026.

{
  "format": "rafael-fuentes-article/1.0",
  "language": "en",
  "title": "Google launches Gemini 4 Hydrogen: fast model for agents at $1.50/M tokens",
  "summary": "Google completes the Gemini 4 family with Hydrogen, a fast model for agents and daily code at $1.50/M tokens, unified 1M window and configurable thinking; warns of more refusals in multimodal.",
  "author": "Rafael Fuentes",
  "publicationDate": "2026-11-13T20:01:16.036+01:00",
  "eventDate": "2026-11-12",
  "url": "https://falifuentes.com/rafael-fuentes-2026-11-13-pm-en/",
  "translation": {
    "language": "es",
    "url": "https://falifuentes.com/rafael-fuentes-2026-11-13-pm-es/"
  },
  "topics": [
    "AI",
    "Artificial Intelligence",
    "Cybersecurity",
    "Cybersecurity News",
    "English"
  ],
  "sources": [
    "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available",
    "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen"
  ]
}