Google ships Gemini4Hydrogen: 2.7x faster than Sonnet5.5 with record-low 8% hallucination rate

AGI Hunt · wechat · 2026-11-13

Google released Gemini4Hydrogen, the third Gemini4 model in six weeks (the long-rumored Flash tier), targeting everyday tasks, coding, and agent workflows — available free.

Benchmarks

ArtificialAnalysis index: 51 (high) / 54 (xhigh), within 1 point of Claude Sonnet5.5 and GPT-6.1Sol peaks, at 32% of Sonnet5.5's per-task cost;
Hallucination rate just 8% vs 43% (Sonnet5.5) — among the lowest AA has recorded;
Weak on terminal agent work: 54% on Terminal-Bench 4.0 vs Sonnet5.5's 64%; also trails on OSWorld.

Speed

7.3s average step in Agent Arena (2.7x faster than Sonnet5.5); 82% of tasks done in 10 minutes; +14.8% net improvement across 6,500 real sessions (rank #5). Outputs 2.8x more tokens than Argon.

Pricing & thinking tiers

Four tiers (low/medium/high/xhigh, xhigh is new): 27/41/51/54 index points at $0.29–$3.16 per task; xhigh doubles tokens for +3 points;
Flat pricing across 1M context: $1.5 in / $7.5 out per M tokens, cached input $0.075.

Orchestrator

Antigravity gains an Orchestrator mode: Carbon plans, Hydrogen codes and tests. Near-Carbon success rates with 24% less quota and 19% faster. Ultra-only for now.

Migration notes: model ID gemini-4-hydrogen; 1M shared context, 64K sync output cap (use background for longer), default thinking high, no explicit caching, temperature/topp ignored.

Safety: agent out-of-bounds behavior down up to 81%, indirect prompt-injection resistance up 2.3x, but over-refusal up 5.57 points; bio/cyber hit Google's frontier safety framework early-warning thresholds.

Original post →

Google ships Gemini4Hydrogen: 2.7x faster than Sonnet5.5 with record-low 8% hallucination rate

More from coding & agent