Google’s Gemini 4 Hydrogen (High) ranks fifth on Arena’s Agent Arena leaderboard with a net improvement score of 14.80%, six places above Gemini 4 Argon (High) at 10.72%, when we read the board at 15:44 UTC on 13 November. In real agent sessions its median task costs $1.62, against $0.71 for Argon.
By our arithmetic, Hydrogen’s median task costs about 2.3 times Argon’s and about 12% more than Gemini 4 Carbon (High)’s $1.45, even though Google’s pricing page lists Hydrogen at $1.50 and $7.50 per million tokens against $2 and $10 for Argon and Carbon. The board shows Hydrogen using a median 69.3K output tokens per task, about 2.9 times Argon’s 23.7K.
Business Pill · Usage-Based Pricing Explained
An evergreen Business Engineer explainer on usage-based pricing, paying for work done rather than for seats, the model behind per-token API rates. It teaches the general idea only and does not discuss Agent Arena or these models.
The key insight: As we read it, Arena’s board shows the other side of Hydrogen’s pricing: it ranks fifth on Agent Arena at 14.80%, six places above Gemini 4 Argon, yet its median task costs $1.62 against Argon’s $0.71, about 2.3 times as much by our arithmetic, on about 2.9 times the output tokens.
Google’s launch post said Hydrogen “ranks #5 on Agent Arena, arena.ai’s live benchmark”, and the board we read shows it fifth. Above it are Claude Fable 5.2 (Max) at 30.48%, Claude Opus 5.5 (Max) at 19.51%, Gemini 4 Carbon (High) at 16.90% and Claude Sonnet 5.5 (Max) at 16.07%.
Hydrogen’s score carries a 95% interval of ±2.23 points, from about 12.6% to 17.0%, which overlaps Carbon’s and sits clear of Argon’s 9.8% to 11.6%, by our arithmetic. The page marks Hydrogen’s entry Preliminary, on 6,465 sessions, about 12% of Argon’s 54,066 and the fewest among the top 11 entries.
Net improvement score with rank, median cost per task and median output tokens per task for Gemini 4 Hydrogen, Carbon and Argon and Gemini 3.8 Flash, all at the High setting, on Arena’s Agent Arena leaderboard (Overall view), read at 15:44 UTC on 13 November 2026. Hydrogen’s entry is marked Preliminary.
Tokens Per Task Drive the Cost
Hydrogen uses a median 69.3K output tokens per task on the board, against 48.1K for Carbon and 23.7K for Argon. That is about 1.4 times Carbon’s and 2.9 times Argon’s, by our arithmetic. As we read it, that is how a model with lower token rates can end up with a higher median cost per task.
It is also quick. Hydrogen’s median task takes 5.6 minutes, the shortest among the top five entries, against 10 minutes for Carbon and 7 for Argon. On steerability, which the board ranks separately, Hydrogen is second at 20.53%, behind Claude Fable 5.2 (Max) at 32.91%.
How Arena Measures It
Arena’s methodology post says that rather than pairwise votes, “rankings are calculated using a methodology we call causal tracing”. It estimates each component’s effect on signals such as task success rates and tool hallucinations, typically against a uniform distribution over components. Our 11 November read of the board had Argon ninth.
As we read the two boards together, per-task cost depends on the test as much as on the model: Artificial Analysis puts Hydrogen below Argon per task, while Arena’s real sessions put it above both Argon and Carbon.
Arena, Agent Arena methodology post, 4 June 2026
“Rather than pairwise votes, rankings are calculated using a methodology we call causal tracing.”
Three Implications
FOR TEAMS CHOOSING AN AGENT MODEL Hydrogen (High) is fifth on Agent Arena at a median $1.62 per task; Carbon (High) is third at $1.45, the board shows.
FOR TEAMS ON ARGON Argon (High) is 11th at 10.72% with a median $0.71 per task, about 44% of Hydrogen’s cost by our arithmetic.
FOR SPEED Hydrogen’s median task takes 5.6 minutes, the shortest among the board’s top five entries.
The framework says the token “is simultaneously the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”
As we read it, the Agent Arena figures separate two of those roles: Hydrogen has the lowest unit price of the three Gemini 4 models, yet its median task uses 69.3K output tokens, the most of the three. The framework predates these models and does not discuss them.
Net improvement score against median (P50) cost per task (log scale) for the top 11 entries and Gemini 3.8 Flash (High) on Arena’s Agent Arena leaderboard (Overall view), read 15:44 UTC, 13 November 2026. Hydrogen’s entry is marked Preliminary.
What Is Not Established
The board updates as sessions arrive, and its page still carried an 11 November date label when we read it on 13 November. Hydrogen’s entry is marked Preliminary and rests on fewer sessions than any other top-11 entry. Median cost and tokens describe Arena’s sessions, not any customer’s bill. We did not test the models.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Arena’s Agent Arena board ranks Gemini 4 Hydrogen (High) fifth at 14.80%, six places above Argon. Its median task costs $1.62 against Argon’s $0.71, about 2.3 times as much by our arithmetic, on about 2.9 times the output tokens.
A note on sourcing. Ranks, scores, intervals, sessions and median cost, time and output tokens per task are from Arena’s Agent Arena leaderboard (Overall view) as we read it at 15:44 UTC on 13 November 2026, and the description of the method from Arena’s methodology post of 4 June. Google’s #5 claim is from its Hydrogen launch post of 12 November and the per-token rates from its Gemini API pricing page; the Artificial Analysis figures are from our own read of that leaderboard the same afternoon.
Ratios, differences and interval ranges are our arithmetic. We did not test the models, and Arena and Google were not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
