Gemini 4 Hydrogen ships faster execution

Google released Gemini 4 Hydrogen November 12 through the Gemini API, claiming coding and agent tasks finish 2.5 times faster than Gemini 3.8 Flash at Hydrogen’s default high reasoning setting.

Google’s “Introducing Gemini 4 Hydrogen” advertises million-token input/output capabilities and flat context-wide rates: $1.50/million input tokens, $7.50/million output. Google DeepMind’s model card specifies a shared 1 million-token input-output window, up to 64K default synchronous output tokens, and March 2026 knowledge cutoff.

Reasoning changes the wait

Artificial Analysis’ Intelligence Index v5.0.1 combines nine evaluations:

| Setting | Score | First answer token, seconds | Weighted minutes/task | |---|---:|---:|---:| | Low | 27 | 5.58 | 1.4 | | Medium | 41 | 11.81 | 3.8 | | High | 51 | 17.37 | 8.3 | | xhigh | 54 | 44.79 | 14.5 |

End-to-end output speeds were 171–178 tokens/second, including reasoning. Google’s self-reported results:

| Evaluation | Hydrogen | Claude Sonnet 5.5 | GPT-6.1 Terra | |---|---:|---:|---:| | Cheating-adjusted Terminal-Bench 4.1 | 57.0% | 63.9% | | | Agents’ Last Exam v1.1 | 71.6% | 60.4% | 69.8% | | AutomationBench-AA | 76.1% | 71.8% | 60.2% |

Google’s launch comparison puts Flash at 18.2% on Terminal-Bench. VentureBeat reports Box’s early-access Complex Work Eval accuracy rose from 74% to 79%; Box CEO Aaron Levie said typical tasks took less than half as long, using about 12% fewer tokens. That isn’t inquiry-conversion evidence.

Token rates aren't task costs

Artificial Analysis estimates $3.16/Index task at xhigh, recording 304 million output tokens versus 110 million at high. VentureBeat reports $1.87/high, $0.80/medium, $0.29/low.

VentureBeat also reports: - Cached input: $0.075/million tokens. - Hydrogen standard rates: twice Flash’s introductory rates, which end December 31; Flash’s scheduled rates begin January 1, 2027. - Batch and Flex: half standard rates; targeted turnaround of 24 hours and 1–15 minutes respectively, Flex serving lower-priority requests. - Google Cloud automatic caching threshold: 4,096 tokens versus Flash’s 6,144.

Its account of Google’s migration guide says Hydrogen defaults to high versus Flash’s medium and rejects thinking_budget and MINIMAL. Google’s deprecation schedule lists no Flash shutdown date.

Test replies, not just throughput

Google claims greater-than-fivefold fewer fabricated claims. Its model card reports unjustified refusals increasing 5.57 percentage points versus Flash, particularly on multimodal prompts. Safety results are automated evaluations, not human evaluation or red teaming.

"The practical takeaway: faster generation is useful only if an accurate, usable reply reaches the prospect sooner."

For a small-team owner/founder, that means a controlled comparison with the current inquiry workflow: measure answer accuracy, response time, cost/inquiry and qualified-lead conversion, then cost/qualified lead. Keep human approval before automatic sending. Benchmarks don’t establish customer gains.

The companion Gemini Decision 1 intake coverage addresses another intake evaluation; Hydrogen still needs testing on the business’s own inquiries.

Gemini 4 Hydrogen ships faster execution—but lead response needs its own test

Reasoning changes the wait

Token rates aren't task costs

Test replies, not just throughput

Sources