E

跑得快 2.5 倍,账单先翻了一倍

Google released Gemini 4 Hydrogen on November 12. It is the third model in the Gemini 4 family, following Carbon and Argon, which focus on deep reasoning. Beyond performance metrics, the naming convention is noteworthy: the long-standing 'Flash' suffix has been dropped in favor of a term from the periodic table.

The official data is neat. In default 'high' thinking mode, task completion is 2.5x faster than Gemini 3.8 Flash; Terminal-Bench 4.1 scores rose from 18.2% to 57%; fabricated statements decreased by over fivefold; Artificial Analysis intelligence index reached 51 points, one point higher than sibling Argon's 50, and up to 54 in the new 'xhigh' mode. Early partner tests showed complex chart and database tasks were 2 to 2.8x faster, with token usage reduced by 12% to 23%.

All these figures come from Google's own press release; independent replication is not yet available.

Now, look at pricing. $1.5 per million input tokens, $7.5 per million output tokens, with uniform pricing within the 1M context window regardless of length. Compared to Gemini 3.8 Flash's current rates of $0.75 and $3.75, this is exactly double.

The key question is when this price gap will disappear. Google's pricing page states clearly: Flash prices are retained until December 31, 2026, rising to $1.5 and $7.5 starting January 1, 2027. This means the same-tier pricing was set to take effect in six weeks; Hydrogen simply accelerates this price hike. Cached input pricing follows suit, rising from $0.075 to $0.15 per million.

Thus, this is not a price increase, but an acceleration. For teams settling by token, switching within the remaining six weeks of this year doubles the unit price immediately; waiting until next January would show no change on the books.

Would it be cheaper if calculated per task? The official stance is that speed and token savings offset the cost. A GitHub example modeled 20 model calls per session, 30k input tokens each, 70% implicit cache hit rate, and 1,500 output tokens, calculating a session cost increase from $0.279 to $0.5265, a rise of 89%. This is a hypothetical scenario, not live traffic data, but it illustrates a point: since cached input pricing didn't change, the increase is slightly less than 100%; lower cache hit rates approach the full double-price impact.

Agent scenarios are truly constrained by this logic. An agent may make dozens or hundreds of calls per task; small token savings and speed gains multiply into significant value. However, if the unit price doubles first, the saved 12-23% cannot cover the difference. Promises of savings require replication; price hikes take effect immediately.

Another detail from official documentation: the model card released with DeepMind notes that unjustified refusal rates are 5.57 percentage points higher than 3.8 Flash. In the two safety assessment columns, bio and cyber risks are marked as 'not reaching critical capability levels,' yet simultaneously flagged as 'reaching warning thresholds,' leaving room for interpretation. Google stated that Hydrogen offers no substantial capability leap over Argon, so full critical capability assessments were skipped, with a complete safety report to follow in a few days.

Market reaction on the day is worth noting. On November 13, $Alphabet(GOOGL.US) closed at $295.77, down 0.98%, having touched an intraday high of $304.97 before falling back below $296. The launch of a default model destined for search, apps, APIs, and office spreadsheets resulted in a stock price spike followed by a pullback.

The center of gravity in model competition is indeed shifting toward cost per task. But in this round, pricing moved first, while efficiency remains ex-post verification. Only after third parties run the same batch of tasks on both Hydrogen and 3.8 Flash will the answer to that 89% projection become clear.

The copyright of this article belongs to the original author/organization.

The views expressed herein are solely those of the author and do not reflect the stance of the platform. The content is intended for investment reference purposes only and shall not be considered as investment advice. Please contact us if you have any questions or suggestions regarding the content services provided by the platform.