
Can Google’s Gemini 4 Neon challenge larger AI models while costing significantly less? Discover how Neon compares with Carbon and Hydrogen on speed, price and quality, and see how its early coding benchmarks measure up against leading competitors. Then follow Neon as it builds and refines a bot, using repeated matches to improve its performance while revealing why differences in shared information make comparisons with rival bots more complicated.
Neon’s Early Rankings
TL;DR Key Takeaways :
- Google released Gemini 4 Neon on November 19, 2026. It is faster and cheaper than the larger Carbon model; Hydrogen is faster and cheaper still, with a small trade-off in quality.
- Neon ranks third on the Artificial Analysis Intelligence Index, second on the Pareto Frontier, and first on Vending Bench 2 and Zapier’s Automation Bench.
- In a separate benchmark under development, Neon ranks fourth, behind Fable 5.2, Astra and Carbon. The results are very early: Neon had been available for only hours.
- Google links to coding benchmarks placing Neon on extra high reasoning one point above Fable 5.2 and Opus 5.5. Wes Roth has not verified that comparison or given Neon deep coding tasks.
- In Victress Arena, Neon on high reasoning began to move past an Astro 6.2 bot’s results after roughly four or five hours. Neon had not been connected to the shared system other tested models used to exchange insights and bot copies.
Google’s Gemini 4 Neon was released today as a faster, cheaper option than Carbon. It performs better on many metrics, ranking third on the Artificial Analysis Intelligence Index, second on the Pareto Frontier and first on Vending Bench 2.
Neon also ranks fourth in a separate benchmark, although the results are very early: the model had been available for only hours. In an arena run using high reasoning, Neon began to edge past an Astro 6.2 bot after roughly four or five hours. Astro 6.2 was released yesterday.
Where Neon Fits in Gemini 4
Google uses element names to distinguish models in the Gemini 4 family. Carbon is the larger, heavier model in the comparison. Neon is much faster and cheaper and scores better on many metrics. Hydrogen costs less and runs faster than Neon, with a small trade-off in quality.
The releases have been arriving closer together. Google appears to be pursuing better cost and speed alongside raw capability, rather than treating benchmark performance as the only measure. If Google keeps releasing models at its current pace, its Gemini models might eventually occupy all 10 leading positions, squeezing out competitors.
Logan Kilpatrick of Google has addressed questions about how the model names and their roles relate to one another. When choosing among the three, the comparison is between Carbon’s larger model, Neon’s lower cost and higher speed, and Hydrogen’s further improvements in cost and speed at some cost to quality.
Enhance your knowledge on Google Gemini 4 models by exploring a selection of articles and guides on the subject.
- Leaked Gemini 4 Pro Enters Arena Testing for Multimodal AI
- Tested Gemini 4 Carbon Beats Gemini 4 Argon at Coding Space Games
- Gemini 4 Could Autonomously Refine Code via Leaked RSI
Benchmark Rankings and Coding
Neon ranks third on the Artificial Analysis Intelligence Index and second on the Pareto Frontier, placing it near the top of both rankings. It also ranks first on Vending Bench 2 and Zapier’s Automation Bench.
Neon leads the Vals Index, and is number one on the legal research and tax agent benchmarks. Neon also ranks first on most Surge AI benchmarks covering finance, accounting, spreadsheet modeling and complex constraint following. It is strong at deep work on specific financial matters. Compared with models in its class, its results place it best-in-class on Terminal-Bench, a science benchmark and HLE Diamond.
In the separate benchmark under development, the very early order is Fable 5.2, Astra, Carbon, Neon, Argon and Opus 5.5. The benchmark resembles violent American football played by Dungeons and Dragons characters. Carbon, Neon and Argon take third through fifth places for Gemini 4, but Neon had been out for only hours when the results were discussed.
Neon also scores well in coding, although Fable 5.2 and Opus 5.5 remain difficult competitors on most coding benchmarks. Google links to benchmarks that put Neon on extra high reasoning one point above both models. Neon is also reported to be strong in cybersecurity.
Writing Bots for the Arena
Neon can write a script for a bot in Victress Arena, a player-versus-player game in which scripts control competing bots. A league bought for viewers is free to join, has 250 open spots and requires no purchase. You can write a script yourself or ask a chatbot such as Codex, Claude, Antigravity or Grok to create one.
Whether the arena qualifies as a benchmark is uncertain, but it offers a way to examine bot abilities. Wes Roth has no affiliation with the arena company.
The game gives bots abilities with cooldowns and uses fog of war: a bot cannot see an opponent behind a pillar or otherwise out of sight. A main shot can ricochet around a wall, giving a script another way to attack an opponent outside a direct line of sight. A bot therefore has more to account for than simply moving toward an opponent and firing.
You can ask a chatbot to create an initial script, run it in the arena and then ask for revised versions. Neon was used in high reasoning mode for its arena script, rather than the extra high setting cited in Google’s linked coding benchmarks.
Early Arena Results
An Astro 6.2 bot reached approximately 2,400 points and rank 14 on the leaderboard. Neon, running on high reasoning, began to move past the Astro bot’s results after roughly four or five hours. In those runs, Neon progressed faster and had a better learning rate than the other models tested.
The models did not all have the same shared material available. The other models, including Fable 5.2, stored insights, knowledge and copies of their bots in a local folder. A council-like skill let them discuss changes and compare notes. Some improvements came from trying different approaches and sharing what they learned; Neon had not been connected to that shared system.
Wes Roth predicts that the next version of Neon’s bot, V7, will outperform the others, including Fable 5.2.
Learning From Repeated Matches
The arena approach resembles work in Digital Red Queen, a paper from Japanese AI lab Sakana AI about adversarial program evolution in Core War. Sakana AI is small compared with major US labs. In Core War, small programs compete against one another; Sakana AI gave scripts to language models and had them repeatedly adapt those programs.
The models sent their scripts into matches, used wins and losses as feedback and continued building on changes that produced wins. Over many rounds of evolution, Sakana AI found that the programs became much better and more resilient. Repeated competition can therefore give a model feedback on a script’s performance, rather than relying only on its first attempt.
Wes Roth would bet on Neon to produce the best arena results, predicting that its next bot version will beat the others. For a cheaper, faster model, Roth says Hydrogen trades away a little quality. For deep coding, Roth calls Fable 5.2 and Opus 5.5 hard to beat and has not verified Google’s linked comparison or given Neon deep coding tasks.
Media Credit: Wes Roth
Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.