Gemini 4 Hydrogen Scores 51 on AA Index at $1.87 a Task, Beating Argon's $4.97, but GPT-6.1 Sol Does 52 for $0.76
Google released Gemini 4 Hydrogen on 12 November, the third Gemini 4 tier after Carbon and Argon. It is priced at $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1 million-token context window. Google says it is available in AI Studio, the Gemini API, Gemini for business, the Gemini app and Search, and in Antigravity for paid AI Pro and Ultra subscribers.
Google’s headline claim is that Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, ahead of Gemini 4 Argon at 50, and 54 at a new xhigh setting. Artificial Analysis’ own leaderboard confirms the numbers and adds the part Google left out: what a task costs.
Cost per task, not list price
AA’s leaderboard reports a cost-per-task figure alongside the Intelligence Index.
| Model | AA Index | Cost per task | Time per task |
|---|---|---|---|
| Gemini 4 Hydrogen (xhigh) | 54 | $3.16 | 13.5 min |
| Claude Fable 5.2 (low) | 54 | $3.10 | 6.1 min |
| Gemini 4 Carbon (low) | 54 | $5.22 | 23.9 min |
| GPT-6.1 Sol (xhigh) | 52 | $0.76 | 9.8 min |
| Gemini 4 Hydrogen (high) | 51 | $1.87 | 8.3 min |
| Gemini 4 Argon (high) | 50 | $4.97 | 30.1 min |
| Gemini 3.8 Flash (high) | 31 | $7.77 | 14.8 min |
Against Google’s own lineup, Hydrogen is a clear step. At high it scores one point above Argon for 38% of Argon’s cost and 28% of its time. Against Gemini 3.8 Flash, the model it replaces for most users, it is 20 index points higher at under a quarter of the cost per task, because 3.8 Flash burns far more tokens to get there.
Against OpenAI, the picture is less flattering. GPT-6.1 Sol at xhigh scores 52 for $0.76, about 40% of Hydrogen’s high-setting cost for a higher score. Hydrogen’s low and medium settings (27 and 41, at $0.29 and $0.80) are cheap, but they are not where Google is pitching it.
At xhigh, Hydrogen matches Carbon’s low setting at 54 for $3.16 against $5.22, and finishes in 13.5 minutes against 23.9. Claude Fable 5.2 at low also scores 54, at $3.10 in 6.1 minutes.
Agent Arena: #5, with a cost-per-task catch
Arena added Hydrogen (High) to its Agent leaderboard on 12 November. It ranks fifth with a net improvement of 14.80% (±2.23) over 6,465 sessions, behind Fable 5.2 Max (30.48%), Opus 5.5 Max (19.51%), Carbon High (16.90%) and Sonnet 5.5 Max (16.07%). It is tagged preliminary.
Its median cost per task on Arena is $1.61, higher than Carbon’s $1.44, even though its per-token price is 25% lower. Hydrogen uses 68.8K output tokens per task against Carbon’s 48K. The upside is speed: 5.6 minutes median against Carbon’s 9.9. Bash reliability is weak at 3.76% (±3.21) versus 5.19% for Carbon and 14.42% for Fable 5.2.
On Arena’s Code Arena WebDev board it sits 10th at 1704, behind Carbon at 1768 and Sonnet 5.5 (xHigh) at 1772.
Orchestrator is the real product move
Alongside the model, Google added an “Orchestrator” option in Antigravity that switches between Carbon and Hydrogen inside a single session. Google’s design: Carbon plans in a read-only phase, Hydrogen writes edits and runs tests, and control returns to Carbon when Hydrogen hits friction. It argues static model routers fail because task difficulty changes mid-trajectory.
Google reports near-Carbon task success on Terminal-Bench with 24% less quota consumed and 19% faster completion. “Near-Carbon” is not quantified, and Google’s two posts disagree on the benchmark version: the main blog says Terminal-Bench 4.0, the Antigravity post says 4.1. Orchestrator (Carbon) is limited to Ultra subscribers; Pro subscribers get Orchestrator (Argon).
Other claims worth flagging
- Terminal-Bench 4.1: 57% for Hydrogen against 18.2% for Gemini 3.8 Flash.
- HLE-Diamond 58.8%, Vals Tax Agent Bench 52.3% (Google says #1), Vals Legal Research Bench 55.8% (tied #1), Multimodal Knowledge Agent 53.4%, Agents’ Last Exam v1.1 71.6%.
- Safety: Google says Hydrogen cuts overreach in agentic tasks by up to 81% and misaligned actions by 12x versus 3.8 Flash, and reports a 96.7% score on an internal agent honesty evaluation.
Most of these are Google-run comparisons against “models in its class”, a set Google does not define. The independent data points are the AA Index and Arena, and both say the same thing: Hydrogen is a strong Google-internal upgrade and a competitive, but not dominant, mid-tier option.