Skip to content

llm-gemini plugin adds xhigh reasoning for Gemini 4 models

Short answer

Simon Willison released llm-gemini 0.38.1, a fix that lets the plugin use the xhigh reasoning level on Gemini 4 Hydrogen and Neon. Carbon stops at high. In his test, Hydrogen at xhigh produced far more output tokens and cost much more than at high.

What this means for operators

If your team calls Gemini 4 models through scripts or the llm command line for ticket triage, lead research or report drafting, the reasoning level is now a cost setting you choose per task. Hydrogen at high spent 3.99 cents on one test prompt, and at xhigh it spent 23.61 cents. Keep high as the default for routine runs, and reserve xhigh for the few jobs where a failed answer costs more than the extra spend. Run your own prompts at both levels and compare the output and the bill before you change any default. Carbon cannot take xhigh, so a shared config that sets it for every model will not work there.

Simon Willison has published llm-gemini 0.38.1, a dot release of his plugin that connects the LLM command-line tool to Google's Gemini models.

What changed

The three Gemini 4 models handle reasoning levels differently. Carbon accepts low, medium and high. Hydrogen and Neon accept those three plus a fourth, xhigh. Willison says he missed that when he first added the models to the plugin. The new release corrects it, so xhigh can now be selected for Hydrogen and Neon.

What xhigh costs

Willison also updated his Gemini 4 comparison grid, which runs the same drawing prompt across the models and settings. The Hydrogen figures from that grid:

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

He describes the xhigh result as a significant improvement. The image Hydrogen produced at that level is, in his words, competitive with what Gemini 4 Neon produced. This is a single informal test on one prompt, so it shows the direction of the trade-off and is not a benchmark.

What to take from it

The release is a compatibility fix in a third-party open-source plugin, not a Google announcement. The source does not say anything about Google's own pricing or about changes in the Gemini API. What it does show is that the top reasoning level on Hydrogen multiplies output tokens and spend per call. Anyone who sets a reasoning level in code should know which models accept which levels, and what each level costs on their real prompts.

Practical checks

  • Update the plugin if you use Hydrogen or Neon and want access to xhigh.
  • Do not apply xhigh to Carbon in shared settings, because that model tops out at high.
  • Log output tokens per call before and after a level change, so the cost shows up in your own numbers.

Source: Simon Willison

The tools in this story

  • Gemini - price, who uses it, alternatives
Next step

Visibility Analyzer

An AI news item will not tell you how assistants see your own site. The Visibility Analyzer checks that on your live site: which answers cite you, which pages an engine cannot retrieve, and what to fix first. Free to run.

Run a free visibility audit

Free to run. No card.

Fee
Free
Length
One run, minutes

Free tier: two analyses a day, no card required.