Google Releases Gemini 4 Hydrogen and Dynamic Antigravity Orchestrator

The Three-Question Framework

What Happened?

On November 12, 2026, Google launched Gemini 4 Hydrogen across Google AI Studio, the Gemini API, Gemini for business, Google Search, the Gemini app, and Google Antigravity, alongside a dynamic Orchestrator mode pairing Hydrogen with Gemini 4 Carbon and Argon.

Why It Matters

At $1.50 per million input tokens and $7.50 per million output tokens across its 1M-token window, Hydrogen tops the flagship Gemini 4 Argon score on the Artificial Analysis Intelligence Index (51 at high default, 54 at xhigh vs. Argon's 50) while cutting Terminal-Bench execution times.

What Could Happen Next

Orchestrator (Carbon) and Orchestrator (Argon) are rolling out to Google AI Ultra subscribers while Pro subscribers receive Orchestrator (Argon), with Google planning additional model pairings and plan support in subsequent updates.

Executive Key Takeaways

Gemini 4 Hydrogen exposes four thinking tiers—low, medium, high (default), and xhigh—supporting up to 1 million input tokens and up to 1 million output tokens.

On Terminal-Bench 4.1, Hydrogen scored 57% versus Gemini 3.8 Flash's 18.2% while completing tasks 2.5 times faster at its default high setting and batching file reads natively.

Within Google Antigravity, dynamic Orchestrator handoffs between Carbon and Hydrogen achieved near-Carbon Terminal-Bench success rates while using 24% less quota and finishing 19% faster.

Safety audits reported 99.9% chain-of-thought legibility across over 200,000 audited turns with zero deceptive reasoning edits and under 0.4% false refusals on HealthBench Pro.

Source: Google blog post by Koray Kavukcuoglu and the Google Antigravity Orchestrator release notes, both dated November 12, 2026, alongside official Gemini API pricing documentation.

Positioned as the rapid-execution tier beneath Gemini 4 Carbon and Argon, Hydrogen pairs coding precision with domain agent benchmarks. Google reported 58.8% on HLE-Diamond, a #1 overall score of 52.3% on Vals Tax Agent Bench, a tied #1 score of 55.8% on Vals Legal Research Bench, and a #5 ranking on arena.ai's Agent Arena. On multimodal office work, Hydrogen scored 53.4% on the Multimodal Knowledge Agent (MKA) benchmark over PDFs, charts, audio, and video, and 71.6% on Agents' Last Exam v1.1. On Google's official pricing table, Paid Tier rates sit at $1.50 input, $7.50 output (including thinking tokens), and $0.075 implicit context caching per million tokens, with Free Tier access available in AI Studio.

Rather than relying on a static front-end router that locks a session to one model after the initial prompt, Google Antigravity's Orchestrator hands control back and forth between Carbon and Hydrogen mid-trajectory. In a typical coding workflow, Carbon begins in read-only exploration to inspect dependencies and establish a plan with explicit re-planning triggers before handing execution, package updates, and test runs to Hydrogen. Whenever Hydrogen encounters unexpected test failures or ambiguity, it escalates state back to Carbon to repair the plan. Google added that developers using the Antigravity SDK can wire the two models explicitly for multi-agent issue triage and repository refactors.

Alignment evaluations targeting high-frequency tool use showed steep error reductions relative to Gemini 3.8 Flash. Google stated Hydrogen reduced fabricated claims by more than 5x, lowered agentic overreach by up to 81% and under-service by up to 79%, and cut misaligned actions and task shortcutting across coding evaluations by 12x. The model reached a 96.7% score on Google's internal agent honesty check, became up to 2x less likely to attempt permission-boundary bypasses on impossible tasks, and demonstrated over 2.3x higher resilience against indirect prompt injection on the Gray Swan IPI benchmark.

Limits: Early partner speedups of 2x to 2.8x faster runs and 12% to 23% token reductions over 3.8 Flash come from internal Google enterprise testing rather than independent replication, Gemini 4 Hydrogen does not yet appear on LMSYS Arena's public text Elo leaderboard snapshot, and SWE-bench Verified has not published shared-harness numbers for the model.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations