Gemini 4 Hydrogen on xhigh Rises from 3.99 to 23.61 Cents per Request
The llm-gemini 0.38.1 plugin now supports xhigh, the most "thinking" level, at which a model spends many tokens on reasoning. It is available for Gemini 4 Hydrogen and Neon, but not for Carbon. In the pelican-on-a-bicycle test, Hydrogen on high cost 3.99 cents and took 19 seconds, while on xhigh it cost 23.61 cents and took 2 min 36 s, but its drawing caught up with Neon. That is why a cheap model at maximum settings should be compared with an expensive one at a standard level.
AI-processed from Simon Willison; edited by Hamidun News
Many modern models let you choose how long they "think" before answering. The higher the reasoning level, the more tokens go to reasoning and the more expensive and slower the request becomes. But if the level you need is not enabled in your tool, you can neither test it nor calculate your budget. That is exactly what happened with the Gemini 4 family in the llm-gemini plugin, and Simon Willison released a fix.
What was fixed in llm-gemini 0.38.1
llm-gemini is a plugin for the llm command-line utility. It lets you call Google's models from the command line and from scripts. When the author added three Gemini 4 models to the plugin — Carbon, Neon and Hydrogen — he missed one detail. Carbon offers three reasoning levels: low, medium and high. Hydrogen and Neon have a fourth, highest level, xhigh.
Version 0.38.1 corrects this. The maximum level can now be explicitly requested wherever it is supported. Carbon still does not have it: in the author's comparison table, that cell is marked "Unsupported."
What xhigh actually costs in practice
Willison updated his Gemini 4 comparison table. The task is playful but useful for measurement: draw a pelican riding a bicycle. The input prompt is 40 tokens everywhere, so the price difference comes almost entirely from output tokens, that is, "thinking" and the result itself.
Hydrogen illustrates the point best. On high it used 5,311 output tokens, cost 3.99 cents and took 19 seconds. On xhigh it used 31,477 tokens, cost 23.61 cents and took 2 minutes 36 seconds. Tokens and spending rose roughly sixfold, and time roughly eightfold. In exchange, according to the author's assessment, the pelican reached the level of Gemini 4 Neon. On high, the picture was noticeably simpler: no basket of fish and no sea.
Other figures from the table:
- Neon on high: 9,691 tokens, 14.55 cents, 1 minute 11 seconds.
- Neon on xhigh: 26,546 tokens, 39.83 cents, 3 minutes 47 seconds.
- Carbon on high: 20,305 tokens, 20.31 cents, 4 minutes 11 seconds.
If you divide the price by the number of output tokens, Hydrogen works out to about $7.5 per million tokens, Carbon to about $10, and Neon to about $15. This is a calculation from the table, not an official price list. But it explains the picture: Hydrogen on xhigh costs 23.61 cents, which is cheaper than Neon on xhigh (39.83 cents), and faster by more than a minute.
What follows from this
First: comparing models at the same reasoning level is unfair. Hydrogen on high and Hydrogen on xhigh are, in effect, two different products in terms of price and quality. If you are choosing a model for an agent or code generation, run your task across the full grid of model × level. Count the cost of a successful result, not of a single call.
Second: a cheap model at its maximum level sometimes beats an expensive model at the same level. Here, Hydrogen on xhigh produced a picture comparable to Neon on xhigh and cost less. But the advantage is not absolute: Neon on high cost 14.55 cents and took 1 minute 11 seconds, making it cheaper and faster than Hydrogen on xhigh, though its drawing is simpler. What is better depends on whether you need xhigh-level results.
Third: time. The difference between 19 seconds and nearly three minutes is critical for interactive scenarios. xhigh is unlikely to suit a chatbot, but it is quite workable for a background batch job.
Limitations
This is a single test on one playful task, and the judgment that a result is "on Neon's level" is subjective. A pelican does not prove that the price-to-quality ratio will be the same on your code or data. Prices and timings depend on load and Google's pricing. Even so, the order of magnitude — roughly a sixfold increase in spending when Hydrogen moves from high to xhigh — is a useful benchmark for budgeting in advance.
- Nov 21, 2026Anthropic ends first week on the stock market at a $1.6 trillion valuation, but falls 8.5% in a single day
- Nov 21, 2026Startup Osmo Trains AI on Smells Using 10 Million Human Ratings and Generates Fragrances From Text
- Nov 21, 2026North Korea Shows First Chatbot "Source of Wisdom": Science Only, No Politics
- Nov 21, 2026OpenAI Opens ChatGPT Spaces to Plus Subscribers: Shared Documents with AI
Key Questions Answered
What was fixed in llm-gemini 0.38.1?
llm-gemini is a plugin for the llm command-line utility. It lets you call Google's models from the command line and from scripts. When the author added three Gemini 4 models to the plugin — Carbon, Neon and Hydrogen — he missed one detail. Carbon offers three reasoning levels: low, medium and high. Hydrogen and Neon have a fourth, highest level, xhigh. Version 0.38.1 corrects this. The maximum level can now be explicitly requested wherever it is supported.
What xhigh actually costs in practice?
Willison updated his Gemini 4 comparison table. The task is playful but useful for measurement: draw a pelican riding a bicycle. The input prompt is 40 tokens everywhere, so the price difference comes almost entirely from output tokens, that is, "thinking" and the result itself. Hydrogen illustrates the point best. On high it used 5,311 output tokens, cost 3.99 cents and took 19 seconds. On xhigh it used 31,477 tokens, cost 23.61 cents and took 2 minutes 36 seconds.
What follows from this?
First: comparing models at the same reasoning level is unfair. Hydrogen on high and Hydrogen on xhigh are, in effect, two different products in terms of price and quality. If you are choosing a model for an agent or code generation, run your task across the full grid of model × level. Count the cost of a successful result, not of a single call. Second: a cheap model at its maximum level sometimes beats an expensive model at the same level.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.