Gemini 4 Neon: Google's $3/$15 middle model leads Opus 5.5 and GPT-6.2 Sol on most of its own benchmarks
Google released Gemini 4 Neon on November 18, a 1M-context model priced at $3 input and $15 output per million tokens that slots between Hydrogen and Carbon. The pricing, the benchmark table, what it changes for developers, and the safety caveats.
Google released Gemini 4 Neon on November 18, the fourth model in the Gemini 4 family and the one meant to sit in the middle of the family: deeper than Gemini 4 Hydrogen, cheaper than Gemini 4 Carbon. The API model ID is gemini-4-neon, the price is $3 per million input tokens and $15 per million output tokens with flat pricing across a 1,048,576-token context window, and Google’s own model card shows it ahead of Claude Opus 5.5 and GPT-6.2 Sol on 16 of 19 listed benchmarks.
The pitch is cost per task rather than raw capability. Google says Neon completes tasks at 32% lower cost than Carbon at the same “high” thinking setting, and that its new “xhigh” setting is more than 36% cheaper per task than competing models at their own highest settings. Those are Google’s numbers and should be read as such until independent evaluators publish their own.
Where Neon fits, and what it costs
The Gemini 4 lineup now has four tiers, with Neon positioned as the “deeper analysis” option:
- Hydrogen (November 12): $1.50 / $7.50, the fast everyday default.
- Neon (November 18): $3 / $15, built for “complex coding, analytical reasoning, and production-grade enterprise workflows,” per the model card.
- Carbon (November 10) and Argon (October 13): $2 / $10 through November 30, rising to $4 / $20 on December 1, according to Google’s pricing page.
That December 1 change is what makes Neon interesting for anyone running Carbon in production. Today Neon is more expensive than Carbon per token; in 12 days it becomes 25% cheaper, as 9to5Google points out. Teams that adopted Carbon at its launch price have a concrete reason to benchmark Neon on their own workloads this month.
The rest of the price sheet: cached input is $0.15 (implicit caching only, a 95% discount), Batch and Flex run at half price ($1.50 / $7.50), and Priority costs $5.40 / $27. There is no free tier for Neon. Output tokens include thinking tokens, which matters because the default thinking level is high.
The benchmark table
Google’s model card compares Neon only with Claude Opus 5.5 and GPT-6.2 Sol, using results “as of November, 2026.” Selected rows (Neon, Opus 5.5, Sol):
| Benchmark | Neon | Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|
| Artificial Analysis Intelligence Index v5.0.1 | 61 | 59 | 57 |
| Vals Index v2.2 | 70.9% | 70.1% | 66.1% |
| Terminal-Bench 4.1 | 65.2% | 65.8% | 61.5% |
| FrontierCode 2.0 | 46.2% | 47.3% | 43.8% |
| CursorBench 4.1 | 65.4% | 58.1% | n/a |
| AutomationBench | 56.3% | 42.5% | 41.3% |
| Agents’ Last Exam v1.1 | 75.8% | 71.6% | 73.0% |
| GraphWalks 256k-1M BFS (F1) | 95.2% | 66.8% | 80.1% |
| Chartography | 73.5% | 66.3% | 75.1% |
Three details are worth more than the headline wins. First, Neon trails Opus 5.5 on the two most widely watched agentic coding benchmarks, Terminal-Bench 4.1 and FrontierCode 2.0, by about a point each; Google’s coding lead comes from CursorBench and its own SWE-sweep harness. Second, the SWE-sweep result depends heavily on the harness: 39.0% in Google’s Antigravity harness versus 10.1% for Opus 5.5, but only 12.8% versus 3.6% (and 11.6% for Sol) in the neutral mini-SWE harness. Third, the long-context GraphWalks gap (95.2% against 66.8%) is the largest margin in the table and the most relevant for anyone stuffing whole repositories or document sets into context.
On third-party leaderboards Google cites, Neon ranks third on the Artificial Analysis Intelligence Index (behind larger, slower models), first on Artificial Analysis’s Coding Agent Index at 58.6, first on Andon Labs’ Vending-Bench 2 with a $16,843 result, and second on both the Vals Index and Arena’s Agent Arena (score 1336). Digital Trends notes that independent tests may come out differently, and Notebookcheck points out that Google picked its comparison models.
What changes for developers
Google’s migration notes contain the details that will break code if missed:
- Sampling parameters are gone.
temperature,top_pandtop_kare removed; out-of-range values return a 400. Determinism has to come from prompts and structured outputs instead. - Thinking levels are low, medium, high (default) and xhigh, set via
thinking_level. Theminimallevel used by smaller models is rejected. - Output is 65,536 tokens by default. Long decodes up to the full 1M window require
background: truewith an explicitmax_output_tokens; background mode without it still stops at 65,536. - No explicit caching, no Live API, text output only. Inputs can be text, images, audio, video and PDFs. Function calling, structured outputs, code execution, computer use, file search and Search or Maps grounding are supported.
- Timeouts. Google recommends client timeouts of at least 15 minutes for hard prompts at high thinking levels.
In Google Antigravity, Neon is available to AI Pro and Ultra subscribers across the desktop app, CLI and IDE extensions, and can be paired with Hydrogen in one session: Neon plans, Hydrogen executes. That planner-executor split is the clearest signal of how Google expects the two models to be used together on cost-sensitive agent workloads.
In the Gemini app, Neon takes Argon’s slot in the AI Pro model picker under “Deeper analysis,” and 9to5Google reports Gemini 3.1 Pro is leaving the app. In the API, Argon stays available with no shutdown date announced. Free and AI Plus users do not get Neon.
Safety notes
The model card says Neon was adapted from Carbon, has a March 2026 knowledge cutoff, and showed a capability increase over Argon that stayed below the threshold requiring a full critical-capability evaluation under Google’s April 2026 Frontier Safety Framework. Argon’s determinations therefore carry forward: Google says it “cannot be ruled out” that the model reaches the critical capability levels for biological uplift and for cyber uplift. The full Neon safety and security report is listed as “coming soon.”
Google’s agentic safety claims are strong but self-reported: a 0.0% cheating rate on Terminal-Bench 4.0, a 0.09% misaligned-action rate, and no successful attacks in Gray Swan’s indirect prompt injection test at the highest thinking level with safeguards on. Access to Neon for vetted cyber defenders through the Fairwind program is rolling out “in the next few days.”
What to watch
- Independent evaluations. Terminal-Bench and FrontierCode results from outside labs will show whether the coding story holds without Google’s harness.
- December 1. Once Carbon and Argon double in price, watch how much API traffic moves to Neon, and whether Google adjusts Neon’s own pricing.
- The Neon safety report, particularly whether the full evaluation changes any of the inherited “cannot be ruled out” findings.
- Rate limits. Neon-specific RPM and TPM numbers are not yet in the public docs; production teams will need them before migrating.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Deploy Enterprise MCP with Cloudflare Workers
Learn to secure and scale Model Context Protocol deployments using Cloudflare’s reference architecture for remote MCP servers and centralized portals.
Gemini 4 Hydrogen replaces Flash at $1.50/$7.50: faster agents, triple the Terminal-Bench score, and a higher bill per token
Google released Gemini 4 Hydrogen on November 12 as its everyday model for coding and agents. It scores 51 on the Artificial Analysis index, beats Argon, costs twice Flash's current price, and its model card admits more unjustified refusals.
Gemini 4 Carbon nine days on: top of the index, a third of Claude's task cost, and still no API
Google's October 28 frontier model leads Artificial Analysis at 59 and costs about $16 per GDPx task against Opus 5.5's $51. Developers still cannot call it, its hallucination rate rose on one test, and its linked safety report reportedly covers Argon.
Gemini 4 Carbon reaches the API: $2/$10 until November 30, then double, and a safety report that calls it no material change
Google opened Gemini 4 Carbon to paying developers and AI Ultra subscribers on November 10, with a 1M-token window and a price rise to $4/$20 on November 30. Its safety report carries Argon's risk findings forward, while showing a 12.8-point jump in exploit success.
Gemini 4 Argon Lands With 1M-Token Output and a Safety-Gated Rollout
Google announced Gemini 4 Argon on September 30 with a 1 million token output limit, a DeepSWE state of the art at 77.9%, and defensive-cyber focus, but most users cannot touch it yet: rollout runs through a trusted-defenders program first.