Google's $1.87 AI Task Is the Win. The Router Is Still a Bet

Hydrogen makes Google's default model cheaper to run on an independent benchmark. Its more ambitious promise depends on who controls the thinking dial.

Artificial Analysis figures below are as listed on 14 November 2026.

One thinking setting adds $1.29 to a benchmark task: $3.16 at xhigh minus $1.87 at high. Same model. Same published token tariff. The extra expense comes from asking Gemini 4 Hydrogen to think harder, with thinking tokens billed as output. You might reasonably ask why cheaper intelligence comes equipped with a dial that makes each job dearer.

Google's achievement is the ordinary setting. Hydrogen high is faster and cheaper per independent benchmark task than Argon, with a slightly higher intelligence score. The grander ambition is to save more by routing between Hydrogen and Carbon, a benefit Google has measured only in its own tests. The Reasoning Toll is the extra per-task charge when a higher effort setting generates more billable tokens at an unchanged tariff. That toll matters because you buy work, whereas Google sells units of computation. A purchasing decision has quietly become a software setting.

The Google price schedule asks $1.50 per million Hydrogen input tokens and $7.50 per million output tokens, including thinking. Output costs five times as much, $7.50 / $1.50 = 5. High is the default; Google's documentation warns that xhigh uses more tokens and time. Hydrogen is adapted from Carbon, according to its 12 November 2026 model card. You buy a choice of effort along with the model.

Artificial Analysis lists Hydrogen high at 51 Intelligence Index points, $1.87 per task and 8.1 minutes. Argon registered 50 points, $4.97 and 29.9 minutes on that index. Calculated here from those two figures: ($4.97 - $1.87) / $4.97 = 62.4% per task, and 29.9 - 8.1 = 21.8 fewer minutes waiting. You get slightly more measured intelligence, a much smaller bill and a shorter wait. An Argon customer has reason to try Hydrogen.

Alphabet's filing for the quarter ended 30 September 2026 records Google Cloud revenue of $30,709 million and operating income of $11,121 million, a margin of $11,121 million / $30,709 million = 36.2%. At its 28 October 2026 earnings call, Alphabet gave $42.3 billion of operating cash flow and $51.4 billion in capital expenditure, leaving $42.3 billion - $51.4 billion = negative $9.1 billion of free cash flow. The revenue and operating income are Google Cloud segment figures; the operating cash flow, capital expenditure and resulting free cash flow are Alphabet-wide. In a scenario with Argon's benchmark token mix unchanged after its scheduled 1 December 2026 rates go from $2/$10 to $4/$20, $4.97 × ($4 / $2) = $9.94 per task. Google controls the tariff. You receive the invoice.

Xhigh is where the cheap model starts acting expensive. Moving Hydrogen from high to xhigh raised its score from 51 to 54, or 54 - 51 = 3 points. Output expanded from 110 million to 304 million tokens, or 304 / 110 = 2.76 times as many. The bill climbed from $1.87 to $3.16, an increase of ($3.16 - $1.87) / $1.87 = 69.0%. Elapsed time grew from 8.1 to 14.3 minutes. The Reasoning Toll is optional, and it is paid every time the setting calls for it.

Sol punctures the price card. On Artificial Analysis, OpenAI's GPT-6.1 Sol at max effort scored 55 and cost $1.27 per task. Hydrogen xhigh scored 54 and cost $3.16. That's $3.16 / $1.27 = 2.49 times the bill for one fewer point, despite Hydrogen's $1.50/$7.50 token rates against Sol's $2/$10. Sol generated 63.7 million output tokens against Hydrogen's 304 million, or 304 / 63.7 = 4.77 times as many for Hydrogen. Artificial Analysis lists Hydrogen xhigh at 304M output tokens from its Intelligence Index run. The discount buys more output at a higher total price. The discount is not the bill.

Google's 12 November 2026 launch statement makes the strongest opposing case. Its Antigravity Orchestrator, Google says, "reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster." That is a company benchmark claim. Artificial Analysis provides the independent evidence for the default: Hydrogen high scored 51 against Argon's 50 at $1.87 rather than $4.97. A default that already wins on cost and time gives a router something worth deploying. You might value speed enough to pay more per task, too.

There is a measured counterexample. In FrugalGPT, dated 9 May 2023, Chen, Zaharia and Zou compared a single GPT-4 model costing $33.1 with a FrugalGPT approach costing $0.6 on the HEADLINES dataset at matched accuracy; their table lists 98.3% savings. Selecting how to spend on inference has produced a cheaper measured result before. The comparison stops at 2023 models on one dataset, whereas Google's routing numbers come from its own tests of another system. FrugalGPT measures neither Hydrogen's customer bills nor its retries. Sol still buys 55 points for $1.27 against Hydrogen xhigh's 54 for $3.16 on the cross-provider index.

Artificial Analysis calculates a weighted benchmark cost at list prices, while customers pay for accepted work; its Hydrogen xhigh and Sol max settings also differ. Hydrogen-specific revenue, margin, migration, retry rates and independently measured routing savings remain absent from the cited results. And time has a price the benchmark's dollar column leaves open. Hydrogen high took 8.1 minutes against Sol's 23.5, a reduction calculated here: (23.5 - 8.1) / 23.5 = 65.5%, while costing $1.87 instead of $1.27. For latency-sensitive work, you might pay that premium happily. Its value depends on the buyer.

The pricing calendar makes the incentive clearer. At the published rates, Hydrogen's $1.50/$7.50 is 25% below Argon and Carbon's $2/$10, since $1.50 / $2 = $7.50 / $10 = 0.75. From 1 December 2026, Argon and Carbon are scheduled at $4/$20, making Hydrogen 62.5% cheaper because $1.50 / $4 = $7.50 / $20 = 0.375. Gemini 3.8 Flash costs $0.75/$3.75 per million input/output tokens through 31 December 2026, so Hydrogen's $1.50/$7.50 is twice Flash's tariff. From 1 January 2027, Flash is $1.50/$7.50, equal to Hydrogen. Raising the alternatives widens the advertised discount without changing Hydrogen's own tariff. You are buying a hierarchy of effort with pricing that Google controls.

I would buy Hydrogen high where it passes my own quality checks and leave the router's claimed gains out of any valuation premium. For a CTO, paying for a completed task is the relevant contract with reality. For an investor, the measured default improvement is substance; routing remains upside. By the end of January 2027, my thesis that the price card does not predict the bill would be wrong if Artificial Analysis lists Hydrogen xhigh at or below GPT-6.1 Sol in cost per task on the same index. Until that happens, pay for the cheaper default against Argon and demand independent results before paying for the grander story.