Google's 25% list-price Hydrogen discount is the other half of a doubled Carbon list price

Hydrogen can shrink the bill for work Google routes away from Carbon. It also gives Google a way to raise the price of the work that stays.

Google is discounting the AI worker while doubling the price of its boss. Gemini 4 Hydrogen launched on November 12 at a 25% list-price discount to Gemini 4 Carbon's November prices. Carbon's list price doubles on December 1. CTOL reads Google's wager as routing: choosing the model for each task keeps buyers inside its system as premium prices rise. Customers may get cheaper work and a higher bill.

On November 11, Business Insider's Hugh Langley reported internal staff access. He relayed an engineer's assessment of Hydrogen as "very capable, fast and efficient". Google launched Hydrogen the next day. Google described its purpose: "Hydrogen is built for rapid execution across diverse agentic workflows." In Antigravity, Google's Orchestrator switches models within a session. Carbon can coordinate while Hydrogen executes. Google writes the rules and the prices.

The discount gets larger as the premium price doubles

Google's pricing page lists Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens. Carbon and Argon cost $2.00 and $10.00 through November 30, then $4.00 and $20.00 starting December 1. Hydrogen has no scheduled increase on the page.

On CTOL's arithmetic, assuming three input tokens per output token, Hydrogen blends to $3.00 per million tokens at list price. Carbon blends to $4.00 through November 30 and $8.00 from December 1. That widens Hydrogen's discount from 25% to 62.5%.

The buyer who stays on Carbon pays twice as much for unchanged token usage. The buyer who offloads work has a larger discount against Carbon. The denominator is doing some of the selling.

Routing can save a customer money and still raise the bill

On CTOL's arithmetic, moving half of a million-token Carbon session to Hydrogen in November cuts the bill from $4.00 to $3.50, before verification and retries. After December 1, the same split costs $5.50. That saves $2.50 against an all-Carbon $8.00 bill, yet costs $1.50 more than November's all-Carbon $4.00.

A customer able to shift half the tokens keeps $2.50 of the December all-Carbon charge. A customer stuck on Carbon faces an $8.00 bill where November required $4.00. On CTOL's reading of list prices and Cloud margin, Google earns less per token routed to Hydrogen and more on the Carbon work that remains. Hydrogen's own margin is undisclosed, so who keeps the saving is not established.

Google reports near-Carbon success in Antigravity on Terminal-Bench 4.0, with 24% less quota, an allocated usage allowance, and 19% faster completion. Quota is not an invoice. The missing measure is cost per accepted task, including the cost of model selection, checking answers and retrying failures. A wrong routing choice spends tokens on work that may have to be done again.

Even "cheaper" depends on the comparison. Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026; Hydrogen charges twice as much. Google says unnamed early partners ran 2x to 2.8x faster than Flash with 12% to 23% fewer tokens. It reports a Terminal-Bench 4.1 score of 57.0% against Flash's 18.2%. The model card also records unjustified refusals 5.57pp higher than Flash. A refusal that blocks valid work has a cost.

Caching or a single capable model can beat Google's router

Google's own pricing creates an awkward counterargument. Its pricing page charges $0.075 per million tokens for Hydrogen's implicit cached input, repeated context reused instead of processed afresh, against $1.50 uncached. When context repeats, caching can dwarf the model discount.

DeepMind's model card also puts Claude Sonnet 5.5 ahead of Hydrogen on cheating-adjusted Terminal-Bench 4.1, 63.9% to 57.0%. It leads on the OSWorld-2.1 offline subset, 83.9% to 77.6%, and on Terminal-Bench-Science 0.1, 59.9% to 51.4%. Anthropic's pricing page lists Sonnet 5.5 at $2.00 input and $10.00 output per million tokens, and Opus 5.5 at $4.00 and $20.00, the prices Carbon reaches on December 1. Buyers whose work resembles those tests could pick one model and avoid routing overhead and failed cheap attempts.

The objection matters even though Google's table awards Hydrogen seven wins out of ten against Sonnet. Google publishes no cache-hit rate or fully loaded cost per accepted task across these alternatives. Repeated context favors caching. Unfamiliar tasks can make a routing choice valuable. Google's Terminal-Bench 4.0 quota reduction supports that possibility, while leaving open whether a single model would be cheaper on particular workloads. Neither benchmark table establishes who pockets the saving.

The seller can own the allocator without keeping its profit

AWS S3 Intelligent-Tiering, launched November 26, 2018, moved storage objects to an infrequent access tier after 30 consecutive days without access. AWS owned the allocation rule. Storage access is observable; task difficulty must be predicted before an agent runs.

Google's Performance Max, launched November 2, 2021, automated ad allocation. Google reported an average 13% more incremental conversions at a similar cost per action, from its own July to September 2021 data. In both cases the seller supplied the allocator, and for Performance Max Google also supplied the measurement. Neither proves who kept the profit.

Cloud margin is also in play. Google Cloud reported $30.709 billion in Q3 2026 revenue and $11.121 billion of operating income, a 36.2% margin versus 23.7% a year earlier. Alphabet chief financial officer Anat Ashkenazi said on the earnings call, "We expect margin headwinds in the near term as we utilize this capacity," referring to third-party compute. Alphabet shares were up 2.07% intraday at 3:04 PM EST on November 12, after rising 2.11% on November 11, the day of the leak. CTOL does not attribute either move to Hydrogen.

Google's costs of serving Carbon and Hydrogen remain undisclosed. Higher revenue might cover higher capacity costs; developers might keep the discount or pass it to their customers. A quota chart cannot decide.

On December 1, 2026, check whether Carbon and Argon rise as scheduled to $4.00 input and $20.00 output per million tokens, or remain at $2.00 and $10.00. Google controls its list prices. Reserve judgment on who profits until the cost per accepted task is measured.