← November 19, 2026 briefing · added in the 16:00 KST update

Gemini 4 Neon follow-up: analysis projects Agent Duel Elo of 1,066, near Opus 5.5's 1,078; ~2.8x Carbon's speed on OpenRouter, pricing more confusing (community)

Announced November 18, 2026

Verified (fact)

Two community analyses appeared right after the Gemini 4 Neon release. An analysis on a personal blog (micjm.com) assumed Neon delivers the same performance as Gemini 4 Carbon, only with faster responses. Under that assumption, it calculated Neon's projected Agent Duel Elo at 1,066, compared with 1,078 for Opus 5.5. On launch day, OpenRouter served Neon at 127 tokens per second and Carbon at 46 tokens per second. The win-rate prediction used model identity, response speed, and whether votes were cast early as variables. The analysis received 209 points on Hacker News. Separately, Simon Willison, while adding Neon support in llm-gemini 0.38, noted that Gemini 4 pricing has become more confusing. The existing lineup is Hydrogen, Argon, and Carbon; Argon and Carbon are priced the same, and prices will rise when the initial discount ends on December 1. Yet the Neon announcement described the lineup without Argon, he said.

Why it matters (interpretation)

Carbon has been seen as slow for its performance level. The analysis suggests that if Neon solves the speed problem, it could compete with Anthropic's top model on agentic tasks. However, the Elo figure is an estimate built on the assumption that its performance matches Carbon's. With the December 1 price change approaching, actual per-model pricing also needs to be checked separately.

Confidence low · official source pending

Sources

Story thread

  1. 2026-11-19 Google DeepMind unveils Gemini 4 Neon, generally available in GitHub Copilot on launch day
  2. 2026-11-19 Gemini 4 Neon follow-up: analysis projects Agent Duel Elo of 1,066, near Opus 5.5's 1,078; ~2.8x Carbon's speed on OpenRouter, pricing more confusing (community)
  3. 2026-11-20 Gemini 4 Neon follow-up: tops engineering benchmark 'Models Engineering Index' at 68.0, ahead of Fable 5.2 (65.3) and GPT-6.2 Astra (61.5) (per community posts)

More about Google