Google says Hydrogen is 2.5× faster than 3.8 Flash. All Agent Encyclopedia's November 13 cross-check argues that "faster" needs unpacking: what's fast is generation throughput, not response speed. Artificial Analysis's independent measurement tells a different story: 41.11 seconds to first token at xhigh, and 13.0 minutes to complete the same index run against a 12.9-minute median. This article puts the official claims next to the independent measurements and works out where the 2.5× actually lives — and where it doesn't.
Hydrogen wins only one of them:
| Metric | Hydrogen (xhigh) | Peer median | Notes |
|---|---|---|---|
| Generation throughput | ~198 tokens/s | — | Once it starts writing, it writes fast — this is what the official "2.5× faster" refers to |
| Time to first token | ~41 seconds | ~13.6 seconds | A long wait before writing starts; AA read 41.11s on Nov 12, 40.55s in the Nov 13 FAQ |
| End-to-end time | ~13.0 min for the AA index run | ~12.9 min | No faster overall — thinking time eats the throughput advantage |
| Total output over the index run | 304M tokens | 72.8M tokens | It thinks long, then writes a lot |
In other words, Hydrogen's working pattern is "think for a long time, then write a great deal in one go". The thinking wait (41 seconds) plus the very long output (304M tokens) brings end-to-end time back to the median. The official 2.5× refers to task completion speed and generation throughput — not how long you wait after pressing enter before the first word appears.
Moving from the default high level to xhigh raises the AA index from 51 to 54 — 3 points — while cost per task rises from $1.87 to $3.16 (+69%) and time per task from ~6.0 minutes to ~13–14 minutes. More importantly: Terminal-Bench 4.0 is 54% at both levels — on terminal tasks, xhigh's extra thinking buys no points at all. Winbuzzer called the results "uneven": scores rise, but task time more than doubles.
That doesn't mean xhigh is useless. On Terminal-Bench-Science 0.1, xhigh rises from high's 41% to 50% — a real gain. But the gain is task-dependent: science reasoning benefits, terminal work doesn't. So the right way to use xhigh is "per task" — measure it on your own tasks and enable it only where it shows a gain; turning it on by default means paying 69% more for 3 points.
All Agent Encyclopedia's judgment: Hydrogen's most buyable advantage isn't speed, it's the hallucination rate. The AA-Omniscience hallucination rate is just 8%, against 43% for Claude Sonnet 5.5 and 52% for GPT-6.1 Sol — a gap measured in multiples, not percentage points. The official line points the same way: fabricated claims down more than 5× versus 3.8 Flash, and the model card's safety evaluation shows image-to-text safety up 7.00 percentage points.
For agents, fewer hallucinations matter more than speed: agent errors cascade — one invented API parameter can invalidate the next ten steps, and rework costs far more than waiting. Hydrogen's role is the "execution tier", and execution is fundamentally about reliability: correct code, correctly read documents, fewer invented facts. The 8% hallucination rate is its biggest upgrade over 3.8 Flash.
But don't look only at the upside: the model card shows unjustified refusals up 5.57 percentage points (worse), especially on multimodal prompts — it's more cautious, at the cost of occasionally refusing what it should answer.
| Scenario | Advice | Why |
|---|---|---|
| Interactive chat, IDE inline completion, anything a user waits on | Don't use Hydrogen | ~41 seconds to first token is too long to wait |
| Everyday coding, single-shot code generation, rapid prototypes | Default high level | The out-of-the-box level: 51 points, $1.87/task; drop to medium (41 points, $0.80) to save money |
| Batch runs, background jobs, parallel workers, the execution role in Orchestrator | Best fit | Nobody waits on the first token, so throughput and low hallucination shine; open xhigh only on tasks where you measured a gain |
Reference source: this article is an original rewrite produced by this site; facts and figures were checked against All Agent Encyclopedia: what's fast is throughput, not response (Nov 13, 2026)
← Back to the Gemini 4 Hydrogen topic page