Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarX: @simonw→ original

xhigh Mode in Gemini 4 Hydrogen Costs 6x More Than high: 23.61 Cents Instead of 3.99 Per Response

Gemini 4 Hydrogen has a setting for how much the model thinks before it answers: high, and an even higher level, xhigh. Simon Willison asked the model to draw a pelican riding a bicycle. The high level used 5,311 tokens (units of text that are billed) and cost 3.99 cents. The xhigh level used 31,477 tokens and cost 23.61 cents. That is nearly six times the price, but according to him, the drawings differ noticeably. So xhigh is worth enabling for hard tasks, not as a default.

AI-processed from X: @simonw; edited by Hamidun News
xhigh Mode in Gemini 4 Hydrogen Costs 6x More Than high: 23.61 Cents Instead of 3.99 Per Response
Source: X: @simonw. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @simonw ↗
Editor-in-Chief
High5,311 tokens and 3.99 cents per drawing
Xhigh31,477 tokens and 23.61 cents, 5.9 times more
Key metric #3The price per token is nearly the same; the bill grows because the reasoning is longer

Modern models can "think" before they answer, and many offer a setting for how long. Gemini 4 Hydrogen has reasoning levels: high and xhigh (even higher). The longer the model reasons, the more text it generates. That text is measured in tokens (small pieces of words), and you pay for them. Simon Willison, a well-known developer and author of many informal tests for language models, showed with one example what this difference costs: xhigh cost 23.61 cents, and high cost 3.99 cents for the same request.

What Was Compared

Willison has long tested models with the same playful task: asking for a pelican riding a bicycle as a vector image. The task looks simple, but it requires assembling a recognizable object from basic shapes, getting the proportions right, and arranging the details. He ran the same model, Gemini 4 Hydrogen, twice: once at the high level and once at xhigh.

Resource results:

  • high: 5,311 output tokens, cost 3.99 cents;
  • xhigh: 31,477 output tokens, cost 23.61 cents.

According to the author, the difference between the two pelicans is "very significant." The images were attached to his post, and the quality assessment is entirely his. We did not compare the images ourselves and are only retelling his conclusion.

Where the Money Goes

If you divide the cost by the number of tokens, both cases come to roughly $7.5 per million output tokens. This is a rough estimate from the published figures, but it shows the main point: the per-token rate is nearly identical across the two levels. The bill rose because the xhigh run wrote 5.9 times more: 31,477 tokens versus 5,311. The cost rose in almost the same proportion, 5.9 times.

For a single drawing, the difference is about 20 cents, which is pocket change. But picture an agent making hundreds of calls per session, or a service handling tens of thousands of requests a day. There, a 5.9x multiplier turns the difference from "tolerable" into "something you have to explain to the CFO." Also, thirty thousand tokens take noticeably longer to generate than five thousand, so response latency will grow too. This is our reasoning, not data from the post, but it follows from the volume of text.

What to Do About It

There are three practical takeaways.

First: don't enable the maximum level by default. Start with high and move to xhigh only where you actually see the quality difference on your own examples: complex multi-step coding tasks or generating intricate structures.

Second: count the cost of the task, not the token. The price list tells you one thing, but your real bill depends on how much the model decided to "think." Measure the average number of output tokens per typical request at each level and multiply by the rate.

Third: set up routing. Send simple requests to lower levels, medium ones to high, and keep xhigh for cases where the first attempt failed or a mistake is expensive.

Limitations

This is one post and one example. The pelican test is subjective and is not a benchmark with a metric, and a single pair of images does not prove that xhigh will deliver the same gain on your tasks. It is also unknown whether xhigh always spends exactly six times as many tokens: the volume of reasoning depends on the complexity of the request. But the arithmetic is useful regardless. Before turning on the maximum, estimate how many times your bill will grow, and check whether the quality gain justifies that cost.

Story

Key Questions Answered

What Was Compared?

Willison has long tested models with the same playful task: asking for a pelican riding a bicycle as a vector image. The task looks simple, but it requires assembling a recognizable object from basic shapes, getting the proportions right, and arranging the details. He ran the same model, Gemini 4 Hydrogen, twice: once at the high level and once at xhigh. Resource results: high: 5,311 output tokens, cost 3.99 cents; xhigh: 31,477 output tokens, cost 23.61 cents.

Where the Money Goes?

If you divide the cost by the number of tokens, both cases come to roughly $7.5 per million output tokens. This is a rough estimate from the published figures, but it shows the main point: the per-token rate is nearly identical across the two levels. The bill rose because the xhigh run wrote 5.9 times more: 31,477 tokens versus 5,311. The cost rose in almost the same proportion, 5.9 times. For a single drawing, the difference is about 20 cents, which is pocket change.

What to Do About It?

There are three practical takeaways. First: don't enable the maximum level by default. Start with high and move to xhigh only where you actually see the quality difference on your own examples: complex multi-step coding tasks or generating intricate structures. Second: count the cost of the task, not the token.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

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.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…