Field noteNov 19, 20265 min read1 source

Google introduced Gemini 4 Neon, adding a model between Carbon and Hydrogen in its Gemini 4 lineup

In one minute

Google introduced Gemini 4 Neon, adding a model between Carbon and Hydrogen in its Gemini 4 lineup. Simon Willison’s llm-gemini 0.38 post compares their stated prices and shows outputs from one image-generation prompt, giving builders a practical, if limited, starting point for weighing cost, speed, and quality.

What changed: llm-gemini 0.38 examines the new Gemini 4 Neon lineup and adds updated pricing and model comparisons, including how Neon fits alongside Carbon and Hydrogen.

What happened

Simon Willison’s announcement of llm-gemini 0.38 discusses Google’s introduction of Gemini 4 Neon and the model lineup’s pricing. Google’s comparison, reproduced in the post, presents three models: Carbon for the hardest long-horizon projects and deep research; Neon for deep reasoning across software engineering, knowledge work, and science; and Hydrogen for speed and efficiency in everyday coding and agentic workflows. The table gives Artificial Analysis Index quality scores of 60 for Carbon, 61 for Neon, and 54 for Hydrogen. It lists input and output prices per million tokens: Carbon at $4 and $20, Neon at $3 and $15, and Hydrogen at $1.50 and $7.50.

The displayed Carbon price is subject to a promotion: $2 per million input tokens and $10 per million output tokens through November 30. Willison notes that Carbon’s regular price is due to double on December 1. That timing makes the apparent price ranking temporary: at the promotional rate, Neon costs more than Carbon, while the published regular prices place Carbon above Neon. Willison also describes the preceding lineup as Hydrogen, Argon, and Carbon, and says Google’s new comparison leaves Argon out.

The post includes Willison’s own small comparison using a prompt to create a flat-style illustration of a pelican riding a bicycle. For Neon, he tried low, medium, and high reasoning levels, then compared the results with Carbon and Hydrogen at the same levels. At high effort, Carbon produced 20,305 output tokens in 4 minutes 11 seconds, priced at 20.31 cents; Neon produced 9,691 tokens in 1 minute 11 seconds, priced at 14.55 cents; and Hydrogen produced 5,311 tokens in 19 seconds, priced at 3.99 cents. The input was 40 tokens in each case. His visual descriptions characterize the results as progressively less detailed from Carbon to Neon to Hydrogen. The post also shows a separate Neon-generated animation prompted in Google Antigravity using extra effort and OpenStreetMap data.

The release is llm-gemini 0.38, linked from Willison’s page. The post does not specify additional access terms for llm-gemini itself; its comparison of Google’s product table says Neon is available through the Gemini API in Google AI Studio (paid), Gemini for business, Google AI Pro and Ultra subscriptions, Gemini app, and Google Antigravity (paid). The table lists Hydrogen across all Gemini API tiers and Carbon through paid API access, among other Google offerings.

What the archive adds

New Runtime’s earlier coverage tracked llm-gemini 0.35 on October 23, then version 0.36 and Gemini 4 Carbon on November 13 and version 0.37 with Gemini 4 Hydrogen on November 13. This is a new model and pricing development in that ongoing tool and model story, rather than a restatement of those releases.

Trend signal

This does not yet fit a pattern New Runtime tracks. The post offers a useful snapshot of model prices and one task comparison, but it does not establish a durable shift in agent economics, interoperability, or operating practice.

Why it matters

For teams choosing a model for coding or agentic work, the headline model ranking is less useful than the combination of price, latency, and output quality for their own tasks. The illustrated prompt makes the trade-off tangible: Hydrogen was much faster and cheaper in this run, while Carbon spent more tokens and time producing a more elaborate image. Neon sits between them in the observed output and timing, but costs more than Carbon during Carbon’s promotion.

The December 1 change will alter that comparison, so teams evaluating costs should distinguish the temporary Carbon rate from its regular price. These figures can inform a test plan, not settle a production choice: a single creative prompt says little about reliability or performance on software tasks. A useful evaluation should hold the task, effort setting, acceptance criteria, latency budget, and token accounting constant, then compare more than one run before a routing policy is changed.

Limits and uncertainty

Willison’s pelican illustration is an informal, single-prompt comparison, not a controlled benchmark of coding, reasoning, or agent performance. The post does not report repeated runs, variance, or an evaluation rubric for image quality. The table’s quality scores are attributed to Artificial Analysis, but the post does not explain their methodology. Prices and the November 30 promotion are reported as displayed in the comparison; actual costs can depend on usage and applicable terms.

The comparison also does not independently verify availability across every listed Google product or subscription. It does not establish whether the observed token counts, elapsed times, or visual differences would repeat under another prompt, a different effort level, or changing service conditions. The creative image task also cannot show how the models behave on a team’s codebase, tool calls, security constraints, or longer agent workflows.

Reader takeaway

Gemini 4 Neon adds another option, but its place in the lineup depends on the task and on whether Carbon’s temporary discount applies. Treat Willison’s example as a prompt to run your own workload-specific comparison, and budget for Carbon’s stated price change on December 1 rather than extrapolating from the promotional rate.

Evidence / sources

  1. 1simonwillison.net/2026/Nov/18/llm-gemini
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicGeminiExplore the gemini topic hub.
  2. 02topicGoogleExplore the google topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract