What happened
Simon Willison’s release announcement identifies llm-gemini 0.37 as adding support for Gemini 4 Hydrogen, which Google launched on November 12, 2026. The package is a plugin for the llm command-line tool. The announcement does not describe a new interface or workflow beyond adding the model to that package; developers already using llm can use the release to access Hydrogen through it.
Willison lists Gemini 4 Hydrogen at $1.50 per million input tokens, $0.075 per million cached input tokens, and $7.50 per million output tokens. He compares those rates with Carbon and Argon, each listed at $2/$0.10/$10 through November 30 and $4/$0.20/$20 from December 1. Willison notes that the two higher models share the same listed price. Google’s pricing description, quoted in the announcement, says Argon delivers Gemini 4 intelligence with faster responses and lower token spend per turn than Carbon.
To illustrate the differences, Willison asked the three models to generate a pelican riding a bicycle at high, medium, and low thinking levels. In the high-effort example, Hydrogen used 5,311 output tokens, cost 3.99 cents, and took 19 seconds. Carbon used 20,305 output tokens, cost 20.31 cents, and took four minutes 11 seconds. The described Hydrogen image was simpler. The medium- and low-effort examples likewise show lower token use, cost, and elapsed time for Hydrogen, alongside less elaborate images.
Google positions Hydrogen, in the description reproduced by Willison, as an execution model for everyday software work or as a high-efficiency engine paired with Carbon or Argon. Willison’s announcement also says Hydrogen is the new free-tier model and will power AI features in Google Search. The release note does not specify further limits or access terms for using Hydrogen through llm-gemini.
What the archive adds
New Runtime’s earlier coverage of llm-gemini 0.35 described the package’s PyPI availability on October 23, 2026. Version 0.37 is a new fact on that known story: it adds a newly launched model. New Runtime also noted Hydrogen’s arrival in Cursor and Vercel AI Gateway on November 12, but those are separate access routes rather than package updates.
Trend signal
This release does not yet fit a pattern New Runtime tracks. It adds a model integration to an existing command-line plugin; the announcement does not establish a broader shift in agent architecture, interoperability, or completed-task economics.
Why now
The timing follows Google’s launch of Gemini 4 Hydrogen on November 12. Willison’s price table makes the immediate comparison unusually relevant: Carbon and Argon are listed at their lower rates only through November 30, then at double those rates from December 1. Hydrogen’s published price stays below both higher-tier models in that table.
Connections to other events
New Runtime’s earlier coverage recorded Gemini 4 Hydrogen becoming available in Cursor and through Vercel AI Gateway on November 12. llm-gemini 0.37 adds another access route, this time for users of Simon Willison’s command-line ecosystem. The common event is model availability, but the products serve different workflows and the release announcement does not claim feature parity among them.
Why it matters
A plugin update can shorten the path from a model launch to a developer’s own evaluation: llm-gemini users can compare Hydrogen without changing their existing command-line tool. The pelican test illustrates a trade-off teams should examine: lower latency and token cost may come with less detailed output. It is not evidence that Hydrogen is the better choice for coding, tests, or verification.
Teams considering Hydrogen for a workflow should run representative prompts, measure successful task completion and review effort, and compare total cost rather than treating a single image example as a benchmark. The stated positioning suggests a plausible division of work—using a cheaper execution model alongside a stronger model for difficult judgment—but that operating model still needs validation against the team’s own tasks and failure modes.
What happens next
- More developers will compare Hydrogen with Carbon and Argon inside existing llm workflows (3 months, 65%). Watch for: additional llm-gemini examples or task-specific measurements.
- The scheduled Carbon and Argon price increase will make Hydrogen’s relative discount more salient (3 months, 70%). Watch for: the December 1 pricing change appearing in Google’s published pricing.
Limits and uncertainty
The pelican comparison is a single illustrative task, not an independent or systematic evaluation of coding ability. Willison’s announcement does not report results on software-engineering tasks, reliability, or the quality of Hydrogen’s output under production workloads. Its price table records a scheduled December 1 increase for Carbon and Argon, but those future rates are not yet observed in the announcement.
The announcement also does not quantify free-tier or Google Search usage limits, establish how much token spend Argon saves per completed task, or compare the models on a shared coding benchmark. The elapsed times and image descriptions are useful demonstrations of one prompt and selected thinking settings, not a general latency or quality guarantee.
Skill to build now
Create a small, repeatable evaluation set before changing a model route. Include representative coding or tool-use tasks, a pass condition, a review-time measure, and per-task token and latency logging. Run the same set through Hydrogen and the model it might replace at comparable settings. This turns a published price difference into a decision about completed work, including the human cost of checking weaker or less detailed output.
Reader takeaway
llm-gemini 0.37 makes Gemini 4 Hydrogen available to users of Simon Willison’s llm tool. The example shows a substantial speed and cost difference on image generation, paired with simpler output. Treat that as a reason to test Hydrogen on bounded, representative work—not as proof that its lower price makes it a substitute for Carbon or Argon.