Google’s Gemini 4 Hydrogen, at its Extra High effort setting, scores 57.8% on Cursor’s CursorBench leaderboard, level with Claude Opus 5.5 (Max), when we read the page at 12:08 UTC on 15 November. Cursor puts Hydrogen’s cost per task at $5.86, against $13.43 for Opus 5.5 (Max).
By our arithmetic, that is about 44% of the Opus cost per task, or about 2.3 times less. At its Low setting the same model scores 27.9% at $0.28 per task and ranks 82nd of the page’s 91 entries.
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 and per-task costs. It teaches the general idea only and does not discuss CursorBench or these models.
The key insight: As we read it, CursorBench prices the effort dial: Gemini 4 Hydrogen at Extra High matches Claude Opus 5.5 (Max) at 57.8% for $5.86 a task against $13.43, while at Low it costs $0.28 and scores 27.9%.
Cursor’s page says “We evaluate agents on ambiguous, multi-file tasks from real Cursor sessions”. Hydrogen at Extra High is sixth. Above it are Gemini 4 Carbon at High (60.4%), Medium (60.1%) and Low (58.6%), Claude Fable 5.2 (Max) at 60.0% and Claude Opus 5.5 (Max) at 57.8%.
Against Carbon at High, Hydrogen at Extra High scores 2.6 points lower at about 64% of the cost per task, $5.86 against $9.22, by our arithmetic. Against Gemini 4 Argon at High, 52.9% at $4.57, it scores 4.9 points higher for about 28% more per task.
CursorBench score at the lowest and highest listed effort setting for Gemini 4 Hydrogen, Claude Opus 5.5, Claude Fable 5.2, Gemini 4 Argon and Gemini 4 Carbon, with Cursor’s cost per task for each setting, from Cursor’s CursorBench leaderboard read at 12:08 UTC on 15 November 2026.
What Was Posted on X
Gemini 4 Hydrogen is priced at $1.50/M input and $7.50/M output tokens.
Cursor’s own account, on 12 November, gave Hydrogen’s list price of $1.50 per million input tokens and $7.50 per million output tokens and pointed developers to its comparison of every effort level. The CursorBench figures in this piece come from Cursor’s leaderboard page, read on 15 November.
Google’s small Flash-tier Gemini 4 Hydrogen has a pretty interesting performance curve.
On CursorBench 4.0, at Low effort ($0.28/task), it scores just 27.9%, below Haiku 5.5.
But push the exact same model to xhigh thinking and things change dramatically:
Score jumps to 57.8% (+29.9 points).
Cost rises to ~$5.86/task.
It ties Claude Opus 5.5 Max.
Completes agent tasks in 14.5 minutes vs 45.3 minutes for Fable 5.2 Max.
It also scored 44.6 FCI on Princeton and DeepMind’s new 5-hour FrontierCS 2 benchmark, beating GPT-6.1 Terra, Sol, and Claude Fable 5.1.
An AI commentary account posted the CursorBench numbers on 15 November. Its 27.9% at $0.28 for Low, 57.8% at about $5.86 for xhigh and the tie with Claude Opus 5.5 Max match Cursor’s page as we read it. Its task times and the FrontierCS 2 score are not on the CursorBench page we read, and we do not use them.
The Effort Setting Moves the Score Most
Cursor lists Hydrogen at four settings: Low at 27.9% for $0.28 a task, Medium at 46.7% for $1.76, High at 54.3% for $5.59 and Extra High at 57.8% for $5.86. From Low to Extra High the score rises 29.9 points while cost per task rises about 21 times and tokens per task about 18 times, from 8,975 to 164,610, by our arithmetic.
That is the widest gap between lowest and highest setting among the models in our chart. Opus 5.5 moves 14.1 points from Low to Max, Fable 5.2 moves 6.8, Argon 2.3 and Carbon 1.8. Hydrogen’s last step, High to Extra High, adds 3.5 points for about 5% more per task.
Agent Arena, Two Days On
On Arena’s Agent Arena leaderboard, read at 12:07 UTC on 15 November, Hydrogen at High is still fifth at 14.80% on 6,465 sessions, as in our 13 November read. Its median cost per task now shows $1.53, down from $1.62, about 6% lower by our arithmetic, on 65.3K output tokens against 69.3K.
Carbon’s median task moved to $1.36 from $1.45 and Argon’s to $0.69 from $0.71. The page carried the same 11 November date label on both reads, so the cost columns changed while Hydrogen’s score and session count did not.
What Users Pay on OpenRouter
OpenRouter’s Hydrogen page lists Google’s endpoints at $1.50 per million input tokens and $7.50 per million output tokens, the rates in Google’s launch post. Its weighted average of what customers pay is $0.4685 per million input tokens, about 31% of the list rate, and $7.319 per million output tokens, about 2% below list, by our arithmetic.
The page says “Caching and discounts mean the price actually paid is often well below the listed one”. It shows an 80.6% cache hit rate on the Google Vertex endpoint, which carried 82.0% of Hydrogen’s tokens over one day. Prompt tokens make up about 96% of the 22.0B tokens in its activity panel, by our arithmetic.
As we read the three boards together, Hydrogen’s cost per task depends on the effort setting and the test: Cursor puts it between $0.28 and $5.86, Arena’s sessions at a median $1.53, and OpenRouter shows customers paying about 31% of the list input rate.
Cursor, CursorBench leaderboard note, read 15 November 2026
“Avg cost / task is computed by applying each model’s published per-million-token pricing (input, cache read, cache write, and output) to the tokens it used on each task.”
Three Implications
FOR TEAMS CHOOSING A CODING AGENT Cursor puts Hydrogen at Extra High level with Opus 5.5 (Max) at 57.8%, at $5.86 per task against $13.43.
FOR TEAMS ON ARGON Hydrogen at Extra High scores 4.9 points above Argon at High on CursorBench for about 28% more per task, by our arithmetic.
FOR COST PLANNING On Cursor’s page, Hydrogen’s cost per task runs from $0.28 at Low to $5.86 at Extra High, about 21 times, by our arithmetic.
The guide says a token “is the unit of intelligence consumed; an agent is a process that consumes tokens to produce work. The new job of finance is to manage the ratio between the two — work produced per token burned.”
As we read it, CursorBench shows that ratio moving inside one model: Hydrogen’s tokens per task rise about 18 times from Low to Extra High while its score roughly doubles, from 27.9% to 57.8%. The guide predates Gemini 4 and does not discuss it.
Median (P50) cost per task for Gemini models at the High setting on Arena’s Agent Arena leaderboard (Overall view), our reads at 15:44 UTC on 13 November and 12:07 UTC on 15 November 2026. Hydrogen’s rank, score and session count were unchanged between the reads.
What Is Not Established
Cursor computes cost per task from list prices, and its page notes that “small differences in scores may not be statistically meaningful”. The tie with Opus 5.5 is a tie on Cursor’s tasks, not a general ranking. The CursorBench changelog’s newest entry is dated 7 October, so the page does not date Hydrogen’s rows.
OpenRouter’s figures cover traffic routed through OpenRouter only. Arena’s Hydrogen entry is still marked Preliminary. We did not test the models.
This Week In The AI Supercycle
AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source
This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.
On Cursor’s CursorBench page, Gemini 4 Hydrogen at Extra High scores 57.8%, level with Claude Opus 5.5 (Max), at $5.86 per task against $13.43, about 44% of the cost by our arithmetic. At Low it scores 27.9% for $0.28.
A note on sourcing. Scores, cost, tokens and steps per task and ranks for CursorBench are from Cursor’s CursorBench leaderboard as we read it at 12:08 UTC on 15 November 2026, including its note on how cost per task is computed. Agent Arena ranks, scores, sessions and median cost, time and output tokens per task are from Arena’s leaderboard (Overall view) as we read it at 12:07 UTC on 15 November, compared with our read at 15:44 UTC on 13 November.
Effective prices, cache hit rates, token shares and token counts are from OpenRouter’s Gemini 4 Hydrogen page read at 12:07 UTC on 15 November; list rates are also in Google’s launch post of 12 November. Ratios and differences are our arithmetic. We did not test the models, and Cursor, Arena, OpenRouter 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.
