Gemini 4 Neon processes 127 tokens per second, more than double the speed of Carbon.
*Gemini 4 Neon launches with a three‑fold speed boost over its predecessor. Analysts project its Elo at 1,066, just shy of Opus 5.5’s 1,078, making it a serious contender on the Agent Duel leaderboard.*
Gemini 4 Neon hit the market on November 18, 2026, and immediately rewrote the speed narrative for large language models. Where its sibling, Gemini 4 Carbon, held the top spot on the Agent Duel leaderboard, it lagged at a crawl of 46 tokens per second. Neon bursts onto the scene at 127 tokens per second, a rate that dwarfs Carbon’s and rivals the fastest commercial offerings. The speed surge arrives alongside a modest benchmark gain, prompting analysts to ask whether raw throughput can now eclipse pure model quality. Early data from OpenRouter confirms Neon’s blistering pace, setting the stage for a potential reshuffle of the leaderboard hierarchy.
Gemini 4 Neon debuted on 2026-11-18. The model shows an incremental benchmark improvement over Gemini 4 Carbon while delivering a dramatic speed increase. On OpenRouter the model processed 127 tokens per second, a rate far exceeding Carbon’s 46 tokens per second. The release positions Neon as the fastest Gemini offering to date. Its performance profile suggests it could rival the top‑ranked Opus 5.5, despite sharing benchmark scores with Carbon. The speed jump is the headline change, not a shift in underlying accuracy metrics.
Carbon remains the best model on the Agent Duel leaderboard when early voting is excluded, but its 46 tokens per second rate hampers practical use. Neon’s 127 tokens per second marks a near three‑fold acceleration. The author’s modeling treats speed as a core component alongside model identity and early voting. Adding an interaction term between speed and early voting improved fit, indicating that faster responses can offset voting biases. The analysis suggests speed now carries weight comparable to model identity, reshaping how leaderboard rankings are interpreted.
Projecting Neon’s Elo assumes it matches Carbon’s benchmark quality while running at 127 tokens per second. The calculation yields a projected Elo of 1,066, just below Opus 5.5’s 1,078 in the later snapshot. The 90% confidence interval for the Elo gap spans -94 to +68 points, meaning Neon could either trail or briefly overtake Opus. The win‑rate projection places Neon at 48.3% against Opus, counting ties as half a win. These figures imply Neon is within striking distance of the current leader, though not a guaranteed overtaker.
The leaderboard model incorporates three factors: model identity, speed, and early voting. Initial attempts without interaction performed poorly, prompting the addition of a speed‑early‑voting interaction term. This adjustment succeeded, highlighting that speed influences outcomes almost as much as the model’s intrinsic quality. The author notes that speed’s importance will be explored in a forthcoming blog. The result underscores a shift in competitive dynamics: raw processing speed can now rival algorithmic superiority in determining leaderboard success.
Neon runs at 127 tokens per second, while Carbon processes 46 tokens per second.
Assuming Carbon’s benchmark performance, Neon’s projected Elo is 1,066, just below Opus 5.5’s 1,078.
Neon is projected to win 48.3% of the time against Opus 5.5, counting ties as half a win.
Neon’s projected Elo of 1,066 places it within a hair’s breadth of Opus 5.5, and its win probability of 48.3% signals a credible threat. If speed continues to prove as decisive as the new modeling suggests, the next leaderboard cycle could see a speed‑driven model eclipse traditional accuracy leaders. Stakeholders should watch Neon’s real‑world performance closely; the race for AI supremacy may now be measured in tokens per second as much as in test scores.
Sources: [Hacker News article "Gemini 4 Neon looks like an Opus 5.5 competitor"]