Urgent.News

What's breaking now, across hundreds of outlets.

AI

Gemini 4 Neon Costs More Per Token and Less Per Task

Carbon costs $2 per million input tokens. Neon costs $3 per million input tokens. Same family, same 1M context window, and the cheaper one has the higher price tag. That detail stopped me when Google shipped Gemini 4 Neon on November 18, 2026. The pitch is "frontier-level intelligence at lower cost." The token price went up, not down. Both things are true, which is the whole story. Neon writes…

On November 18, 2026, Google launched Gemini 4 Neon, positioning it as a "frontier-level intelligence at lower cost" compared to its predecessors. However, the pricing structure is more complex than meets the eye. While Neon boasts a lower cost per finished job, it comes with a higher price per token compared to its Carbon predecessor.

This discrepancy has caused confusion among users, as they might find themselves paying more for a model that writes more tokens and thinks harder by default. Neon sits between Hydrogen and Carbon, inheriting the latter's 1M context window and four thinking levels. Despite being billed as a middle-tier model, it aims to handle complex coding and analysis tasks efficiently.

Google offers access to Neon through various platforms, including the Gemini API, AI Studio, Gemini Enterprise, Antigravity, and the Gemini app for Pro and Ultra, but without a free tier. Google's own benchmarks highlight Neon's strengths, particularly in tasks like legal research, tax analysis, and automation, where it outperforms competitors.

However, it falls short in certain tests, such as Code Migration, where it averages fewer turns per task than its higher-priced counterparts. The cost per task for Neon is lower than Carbon's, but it costs more per token. For instance, Neon high costs $3.74 per minute for 14.8 minutes, while Carbon high costs $7.76 per minute for 37.2 minutes.

Despite its faster performance, Neon's verbosity, which can write up to 283 million output tokens, makes it more expensive than models like Opus 5.5, which writes significantly fewer tokens but at a lower cost. In tasks where accuracy matters, such as the Agent Arena Index, Neon's performance varies, sometimes guessing incorrectly or refusing to answer when uncertain.

This behavior, known as "deceptive negotiation," can be seen as a strength in certain scenarios but raises concerns about reliability. Additionally, the naming of Neon, being atomic number 10, has sparked discussions and confusion among users, who find it challenging to map these new model names to their specific use cases.

Rewritten by urgent.news; may contain errors — check the original.

Read the original at dev.to →

Report a problem with this story

More in AI

Sashiko AI Introducing Linux Performance Specialist Review Stage

The Google-developed Sashiko for agentic AI code review initially focused on the Linux kernel continues helping to uncover new bugs and issues as new patches hit the kernel mailing list as well as…

  • Sashiko AI adds performance specialist review stage for Linux
  • Includes second-order side-effect tracing and anti-noise guardrails
  • Expected 11% increase in resource usage and 3.1% faster commit duration

Your AI Agent Wrote Code. Where Does It Run? A Sandbox Runner in Under 50 Lines of Python

Your AI agent just wrote some Python and wants to run it. Where does that code actually execute? For a lot of agent products the honest answer is "on the same server as everything else, with the same…

  • Sandbox runner implemented in under 50 lines of Python
  • Limits CPU usage, memory usage, file size, wall clock time, output size
  • Executes code in separate process group with restricted environment variables

More from Saturday 21 November →