Gemini logo with headline saying Google's Gemini 4 Neon model is aimed at coding agent work

Google Releases Gemini 4 Neon: Frontier Coding Agent Model at $3 per 1M Tokens

Short answer: Gemini 4 Neon is Google’s new mid tier Gemini 4 model, built to sit between Gemini 4 Carbon and Gemini 4 Hydrogen for coding agents and long multistep work. It costs $3 per million input tokens and $15 per million output tokens with a 1M token context window, and it is live now in the paid Gemini API, Google AI Studio and Google Antigravity. The caveat: nearly every headline score comes from Google’s own post or from leaderboards Google picked, so test it on your own repo before you switch.

Key points

  • Pricing: $3 input, $0.15 cached input and $15 output per million tokens. Cached input is a 95% discount.
  • Google says Neon ranks #1 on the Artificial Analysis Coding Agent Index (58.6) and #3 overall on the Artificial Analysis Intelligence Index (61) at its xhigh thinking level.
  • Four thinking levels. At the default high level, Google claims a 32% lower cost per task than Carbon.
  • Coding numbers from Google: 65.2% on Terminal-Bench 4.1, 46.2% on FrontierCode 2.0, and 39.0% on SWE-sweep inside Antigravity versus 12.8% in the plain mini-SWE harness.
  • Early independent speed reading: about 127 tokens per second on OpenRouter, versus about 46 for Carbon.
  • Access: paid API tier, AI Studio, Android Studio, Gemini for business, and Google AI Pro and Ultra in the Gemini app and Antigravity. A cyber variant goes only to Fairwind Program members.

What shipped with Gemini 4 Neon

Google DeepMind’s Koray Kavukcuoglu introduced Neon as “frontier level intelligence with fast end to end task completion.” In the family, Gemini 4 Carbon stays the model for the hardest long horizon projects, Gemini 4 Hydrogen handles fast everyday execution, and Neon is the balance between depth and speed.

The model takes spreadsheets, PDFs, charts, audio and video in one 1M token context and turns them into structured output. Google lists #1 results on Legal Research Bench, Tax Agent Bench, Zapier’s AutomationBench (56.3%) and Andon Labs’ Vending-Bench 2, plus a #2 debut on Arena’s new Agent Arena Index with a score of 1336 across more than 4,200 real agent sessions.

For engineering work, Google says Neon makes fewer redundant file reads, reports progress as it works and handles context compaction reliably in long sessions. It can run alone or act as an orchestrator that directs parallel Hydrogen subagents in Antigravity.

Gemini 4 Neon pricing compared with other builder models

Neon is not the cheap tier. At $3 and $15 per million tokens it costs double Hydrogen’s $1.50 and $7.50. It is well below OpenAI’s GPT-6.2 Astra at $8 and $40, but above GPT-6.2 Sol at $2 and $10.

Google’s pitch is cost per finished task, not cost per token. It claims Neon at high thinking runs roughly 50% to 80% cheaper per task than top competing frontier models on the Vals and Agent Arena indexes, and at least 36% cheaper at its xhigh setting. Those are vendor numbers. Per task cost depends on how many tokens your agent burns, so measure it on your own workload.

The cached input price is the part builders should notice. At $0.15 per million tokens, agents that resend the same repo context or system prompt on every turn get much cheaper, provided your stack actually hits the cache.

What it means for builders

  • Coding agent teams: Neon is the obvious model to trial if Carbon felt too slow. An independent post by micjm, which runs the Agent Duel leaderboard, measured about 127 tokens per second for Neon on OpenRouter against about 46 for Carbon, and projected Neon roughly level with Claude Opus 5.5 on that leaderboard.
  • Harness matters: the jump from 12.8% to 39.0% on SWE-sweep comes from Google’s own Antigravity harness. If you run Neon through Cursor, OpenCode or your own agent loop, expect results closer to the lower number until third parties report otherwise.
  • Security work: Google reports 90.5% on SEC-bench Pro and 70.7% on CWE-bench v2 for the cyber variant, and a 1.98% refusal rate on Terminal-Bench 4.0 for the main model. Most developers will not get the cyber variant soon, since it is limited to Fairwind members.
  • Multi model routing: Neon gives Google stacks a middle option. A reasonable split is Hydrogen for triage and quick edits, Neon for multi file refactors and debugging, and Carbon only for the longest jobs. Compare that with Anthropic’s lineup in our Claude Fable 5.2 breakdown before you commit.

What to watch

  • Independent Terminal-Bench and SWE-style runs outside Antigravity.
  • Whether OpenRouter speeds hold once traffic grows.
  • Rate limits on the paid API tier, which Google did not detail in the launch post.
  • Whether the cyber variant opens beyond Fairwind members.

Sources

Leave a Comment

Your email address will not be published. Required fields are marked *