Gemini 4 Neon matches Claude Opus 5.5 on independent benchmarks at a third to two thirds of the cost per task, while GPT-6.2 Sol scores four to five points lower at under half Neon's cost. Google made gemini-4-neon generally available on the Gemini API on November 18, 2026, at $3 per million input tokens and $15 per million output tokens with a 1M-token context window, per the AI Studio changelog. Artificial Analysis scores it 61 on its Intelligence Index at $6.44 per task, against 59 and $10.04 for Opus 5.5. Run Neon against your current model on your own tasks this month, and check where you can buy it first: Google's Vertex AI model list did not include Neon on November 19.
The two-point lead is small. The cost gap is the part worth acting on, and it depends heavily on which thinking level you run.
What Google Shipped on November 18
Gemini 4 Neon is the fourth Gemini 4 model, and Google positions it between its fast everyday tier and its long-horizon reasoning tier. In the launch post, Koray Kavukcuoglu says Neon is "rolling out today in the Gemini API in Google AI Studio, Gemini for business (via API and model dropdown), for paid Google AI Pro and Ultra subscribers in Google Antigravity, and for Pro and Ultra subscribers in the Gemini app." The same post says Gemini 4 Carbon handles "your hardest, long-horizon projects" and Gemini 4 Hydrogen "handles fast everyday execution."
The changelog lists four thinking levels: low, medium, high (the default) and xhigh. A thinking level is the setting that controls how much reasoning the model does before it answers, and with it how many output tokens you pay for. The models page gives Neon a 1M-token context window and a maximum output of "64K / 1M tokens," and the changelog describes "long decode up to 1M tokens," so read 64K as the normal synchronous ceiling.
Three availability details matter for an enterprise buyer. The pricing page says "Gemini 4 Neon is not available on the Free tier," which Hydrogen has.
The Vertex AI model catalog lists Gemini 4 Hydrogen, Carbon and Argon as generally available, and did not list Neon when we checked on November 19, 2026. If your security review approved Vertex AI and not the Gemini API, you cannot route production traffic to Neon today.
The pricing page also notes "Implicit caching only (explicit caching is not supported)." If your agents rely on explicitly created caches for long system prompts, plan for cache hit rates you do not control.
How Much Does Gemini 4 Neon Cost Next to Opus 5.5 and GPT-6.2 Sol?
On list price per token, Neon sits between GPT-6.2 Sol and Opus 5.5. Here are the standard-tier rates per million tokens, from each vendor's pricing page as read on November 19, 2026:
| Model | Input | Cached input | Output | Batch (in / out) |
|---|---|---|---|---|
| Gemini 4 Neon | $3.00 | $0.15 | $15.00 | $1.50 / $7.50 |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | $2.00 / $10.00 |
| GPT-6.2 Sol (272K or less) | $2.00 | $0.10 | $10.00 | $1.00 / $5.00 |
| Gemini 4 Hydrogen | $1.50 | $0.075 | $7.50 | $0.75 / $3.75 |
Sources: Gemini API pricing, Claude API pricing, OpenAI API pricing. OpenAI charges Sol $4 input and $15 output above 272K input tokens. Google lists no separate long-context rate for Neon, which matters if your workloads routinely send 300K-token contexts.
Google also sells Neon on Flex at the batch rate and on Priority at $5.40 input and $27 output. Our batch, flex and priority breakdown covers when each tier pays off.
Neon also undercuts Google's own reasoning tier after November. The pricing page shows Carbon and Argon at introductory rates of $2 input and $10 output through November 30, then $4 and $20 from December 1, 2026. Neon's $3 and $15 carry no step-up. Google claims Neon at high thinking has "32% lower cost per task than Carbon," per the launch post. That is a vendor claim, and nobody has tested it independently yet.
Does Neon Beat Opus 5.5 on Independent Benchmarks?
Yes, by two points on Artificial Analysis and under one on Vals, and at a much lower cost per task. Cost per task is the total spend to finish one benchmark task, counting input, output and reasoning tokens together. It is the figure that ends up on your invoice, and it moves differently from the token price.
Artificial Analysis records these at each model's headline setting:
| Model (setting) | Intelligence Index | Cost per task | Output tokens for the index | Time per task |
|---|---|---|---|---|
| Gemini 4 Neon (xhigh) | 61 | $6.44 | 283M | 21.5 min |
| Claude Opus 5.5 (max) | 59 | $10.04 | 218M | 25.2 min |
| GPT-6.2 Sol (max) | 57 | $1.01 | 60.9M | 9.1 min |
Opus and Sol figures come from their own Artificial Analysis pages (Opus 5.5, GPT-6.2 Sol).
The Vals Index, version 2.2 updated November 18, tells the same story with a wider cost gap:
| Model | Accuracy | Cost per test | Duration per test |
|---|---|---|---|
| Gemini 4 Neon | 70.87% | $7.40 | 31m 25s |
| Claude Opus 5.5 | 70.14% | $24.77 | 1h 25m |
| GPT-6.2 Sol | 66.06% | $3.18 | 41m 47s |
On those numbers, Neon finishes an Artificial Analysis task for 36% less than Opus 5.5 and a Vals test for 70% less, while scoring slightly higher on both. Vals ranks Neon second of 42, behind Claude Fable 5.2 at 71.31%.
Neon is not frugal with tokens. It generated 283M output tokens to run the Artificial Analysis index, more than Opus's 218M. It costs less because each token is cheaper: at list output prices, that works out to about $4,245 for Neon and $4,360 for Opus, our arithmetic from the published figures. The Vals gap is wider because Opus used 59.46M tokens per test there against Neon's 15.15M.
One discrepancy to note: Google's post calls Neon "#3 overall" on the Artificial Analysis index, while the Artificial Analysis model page ranked it 6th of 95 when we read it on November 19. Rankings on that board move as models are added, so cite the score, not the rank.
The thinking level decides your bill
Artificial Analysis scores Neon at 60 on high, its default, 55 on medium and 44 on low, against 61 on xhigh. The headline 61 is the expensive setting. High gives up one point and should cost less per task, though Artificial Analysis publishes the dollar figure for xhigh only. Low drops 17 points, below the 51 Artificial Analysis gave Hydrogen at high in our Hydrogen analysis, a model at half Neon's token price.
Set Neon's thinking level per task type in your gateway, the way you would set reasoning effort on Claude or OpenAI models. Our Sonnet 5.5 vs Opus 5.5 analysis showed how a matched score at a lower effort setting can reverse which model is cheaper.
Where Opus 5.5 and GPT-6.2 Sol Still Win
Opus 5.5 still leads on two of the coding benchmarks Google chose to publish, and GPT-6.2 Sol wins on price by a margin Neon cannot close. Google's own comparison table puts Opus 5.5 at 65.8% on Terminal-Bench 4.1 against Neon's 65.2%, and at 47.3% on FrontierCode 2.0 against 46.2%. GPT-6.2 Sol scores 75.1% on Chartography against Neon's 73.5%. These are Google-reported numbers, and the gaps run both ways by about a point.
The strongest case for staying put is cost per point. GPT-6.2 Sol finished the Artificial Analysis index at $1.01 per task, about a sixth of Neon's $6.44, and the Vals Index at $3.18 per test, under half of Neon's $7.40. Sol scores four to five points lower on each. For high-volume extraction, triage or first-draft work where a few points of benchmark accuracy do not change the outcome, Sol is the cheaper default. Our GPT-6.2 launch analysis has Sol's full price table and rate-limit caveats.
Neon's clearest lead is long-context retrieval. Google reports 95.2% on GraphWalks at 256K to 1M tokens, against 66.8% for Opus 5.5 and 80.1% for Sol, on the same DeepMind page. If your agents read whole repositories, contract sets or case files in one call, that is the result to verify first.
What This Means for Your Model Roster
Neon belongs in your evaluation queue as an Opus 5.5 replacement for agentic and coding work, and not yet as an automatic switch. Three conditions have to hold before you move traffic:
- Your own tasks show the same cost gap. Benchmarks measure someone else's tasks, and Neon's high token count means a workload with short answers and little reasoning may not see the savings the indexes show.
- You can buy it through a channel your security team approved. Today that means the Gemini API, Gemini for business or Antigravity. Teams on Vertex AI should ask their Google account team for a date.
- Implicit caching gives you an acceptable hit rate. Cached input at $0.15 is 95% off the input price, so a low hit rate on a long, repeated system prompt can erase part of the per-task saving.
Google also launched Neon into consumer and business apps. Online Tech Tips relays a 9to5Google report that Neon replaces Argon and Gemini 3.1 Pro in the Gemini app picker under "Deeper analysis." Google has not confirmed that on its own pages, so check your Gemini Enterprise seat comparison and admin console before telling users which model they are on.
What to Do Next
This Week:
- Pull 50 to 100 real tasks from your Opus 5.5 logs, with the cost and outcome of each, and run them through
gemini-4-neonat high and at xhigh. Record cost per successful task, not cost per token. - Run the same set through GPT-6.2 Sol. If Sol passes your bar, it is the cheaper choice by a wide margin and Neon's lead over Opus does not matter for that workload.
- Ask your Google account team when Neon will be listed in the Vertex AI model catalog, and whether a long-context rate will apply above 200K tokens.
This Month:
- Add Neon as a routed option in your gateway (LiteLLM, OpenRouter or your own), with the thinking level set per task type and a fallback to your current model. Our routing analysis covers what cache switching costs when traffic moves between providers.
- Measure Neon's implicit cache hit rate on your longest system prompts for two weeks before you count the caching discount in a budget.
- If you run Carbon or Argon at introductory pricing, re-price those workloads at $4 and $20 before December 1, and test Neon as the replacement.
The Bottom Line
Gemini 4 Argon and Carbon both started on limited access, with introductory pricing that doubles on December 1. Neon reached general API availability on launch day at a list price with no scheduled step-up, and its independent scores show a lower cost per task than Opus 5.5. On Artificial Analysis, GPT-6.2 Sol still finishes a task for about a sixth of Neon's cost. Price your top three workloads on all three models before your next model-spend review.
Continue Reading
- Gemini 4 Hydrogen Costs $1.87 a Task vs $7.77 for 3.8 Flash
- Google Opens Gemini 4 Carbon to a Subset of Cyber Defenders
- GPT-6.2 Astra Costs 20% Less Than GPT-6 Astra on API Tiers
- Claude Fable 5.2 Requires 30-Day Retention for Most Buyers
- Claude Sonnet 5.5 Costs More Than Opus 5.5 at the Same Score
- Your LLM Calls Can Cost Half as Much on Batch or Flex
