Google charges $3 per million input tokens for Gemini 4 Neon1. VentureBeat reported that this rate runs 50% above OpenAI's GPT-6.2 Sol2, which puts Sol at $2. On sticker price, Neon loses.
Gemini 4 Neon still scored 61 on the Artificial Analysis Intelligence Index at the xhigh thinking level1. Claude Opus 5.5 (Max, Default Fallback) scored 59 on that Artificial Analysis index3, and VentureBeat lists Opus at $4 per million input tokens2. Artificial Analysis puts Neon (Xhigh) at $6.44 per index task and Opus 5.5 (Max, Default Fallback) at $10.04.
Google shipped Neon on November 18, 2026,1 one day after OpenAI announced GPT-6.2 Astra and Sol. The real fight that week was over the middle of the lineup. If your app still sends every request to a flagship by default, your cost model is probably out of date.
Finished tasks set the real price
A rate card shows price per token. Your business pays the finished-task price, which is the cost of one correct result. Three numbers set it, and Neon's launch moved all three.
| Gemini 4 Neon per task3Artificial Analysis · xhigh thinking | $6.44 |
|---|---|
| Gemini 4 Carbon per task3Artificial Analysis · Intelligence Index | $7.76 |
| Claude Opus 5.5 per task3Artificial Analysis · Max, Default Fallback | $10.04 |
| Neon context windowRequesty · no long-prompt surcharge | 1,048,576 |
The first number is the rate per token. Google lists Neon at $3 in and $15 out per million tokens14, with cached input at $0.15, a 95% discount. Requesty reports Google charges that rate across the whole 1,048,576-token5 context window, with no long-prompt surcharge.
The second number is tokens burned per task. Artificial Analysis found Neon uses about 15% more output tokens per task than Gemini 4 Carbon. Neon still came in cheaper per task because its token prices are 25% lower6: $3 and $15 against Carbon's standard $4 and $2014. Artificial Analysis measured $6.44 per Intelligence Index task for Neon (xhigh3) and $7.76 for Carbon, about 17% less.
The third number is success rate. No leaderboard can give you yours. It lives in your own task logs.
The effort setting moves all three numbers. Google claims a 32% lower cost per task than Carbon at Neon's default high thinking level1. Artificial Analysis says Neon at high effort matches Carbon's score at under half the cost per task. Each figure depends on who measured and which dial they turned.
Only the finished-task price, measured on your own workload, settles the question.
Margin math for a mid-tier default
What does one finished task cost you? What do you charge for it? The gap between those two numbers is your gross margin, and the model bill is your cost of goods.
Margin drives everything else. More margin per task makes each customer worth more over their lifetime. That lets you spend more to win the next customer and still come out ahead on the lifetime value to acquisition cost ratio. Model routing is a pricing decision wearing an engineering badge.
Take one task with 100,000 input tokens and 10,000 output tokens. At Google's list price, Neon costs $0.30 for input plus $0.15 for output, so $0.45 total. At VentureBeat's reported Opus 5.5 rates of $4 and $20,2 the same task costs $0.60. At VentureBeat's implied Sol rates of $2 and $10,6 it costs $0.30.
Now add success rates, which here are hypothetical. Say Neon gets the task right 80% of the time and Opus 5.5 gets it right 95%. Divide cost by success rate. Neon lands at about $0.56 per correct result and Opus at about $0.63.
The break-even sits near 71%. Below that, Neon loses to Opus on this task despite a 25% lower token price6. Sol beats Neon's $0.56 if it gets the task right more than about 53% of the time. The honest read is that Neon earns its price only where its success rate is high.
Verbosity is the hidden line item. Requesty, citing Google's pricing docs, reports that thinking tokens bill as output. Artificial Analysis ranks Neon 92nd of 94 models for verbosity. If Neon burns 20,000 output tokens on our sample task, its bill rises to $0.60, level with Opus at 10,000.
Google's own model card shows how thin the coding lead is. On Terminal-Bench 4.1, Claude Opus 5.5 beats Neon 65.8% to 65.2%.7 On FrontierCode 2.0, Claude wins 47.3% to 46.2%. Google reports Neon first on the Artificial Analysis Coding Agent Index at 58.6.1 Edges that small will not decide your margin.
This week's distraction is the top of the leaderboard. Artificial Analysis ranks GPT-6.2 Astra and Claude Fable 5.2 above Neon, and plenty of teams will wire them into every call. What produces your margin is a table of your own tasks, with cost and pass rate beside each one. You rent the model. You own the table.
These numbers can also hide a service business dressed up as software. If a human repairs one task in five, payroll sits inside your cost of goods and the API bill is the small number. We think most agent startups should price their human review hours before they argue about token rates. The work is dull: a spreadsheet of fifty tasks, rerun after every launch.
Doing it again and again is the edge.
Will this default survive into 2027?
Artificial Analysis counts Neon as Google's fourth model in seven weeks, or one release every 1.75 weeks. OpenAI shipped two models the day before. At Google's pace alone, that is about 30 models a year, so a default picked on November 18 may be stale before January.
Prices move just as fast. The Decoder reports that Carbon's introductory rates of $2 and $10 double to $4 and $20 on December 1.6 Requesty spells out the effect: Neon costs 50% more per token than Carbon through November 30,8 then 25% less. One date on the calendar flips the ranking.
Over five years, the pattern matters more than any one model. If anything near this cadence holds, the model becomes the least durable part of your stack and your measurement harness the most durable. The risk is lopsided. Rerunning 50 tasks on Neon and Opus at our sample rates costs about $52.50, while a stale default leaks margin on every request.
Nobody can say yet whether the middle tier stays the main battleground. A rival could cut flagship prices or ship a cheaper specialist. Neon leads Opus 5.5 by two points on the Artificial Analysis index, and one release could erase that. Our read is that the tier matters less than the habit of re-pricing every few weeks, because that habit catches each price cut first.
Price fifty real tasks before Friday
The goal is one spreadsheet that names the model to call for each task type. Budget about $53 in API spend at our sample rates, plus one afternoon.
Step one: pull 50 real tasks from last month's logs. Mix routine tasks with the ones that broke. Write down what a correct answer looks like for each, so a script or a person can mark pass or fail.
Step two: run every task through Neon and your current default. Neon's API model ID is gemini-4-neon. Log input tokens, cached tokens and output tokens for every call. A cached token is input the API already holds from an earlier identical prompt opening, billed on Neon at $0.15 per million1.
Step three: move your stable system prompt and reference docs to the front of every request. Requesty reports that Google's implicit caching applies automatically to shared prefixes of at least 4,096 tokens. A prefix is the opening chunk of a prompt that stays identical across calls. If your cached count reads zero, something in that opening chunk changes each time, often a timestamp.
Step four: compute cost per correct result. Divide each model's total spend by the number of tasks it passed. Thinking tokens bill as output, so check the output column first when a number looks high.
Step five: write the routing rule. Send each task type to the cheapest model that clears your pass bar. Escalate to a flagship only where one failure costs more than the price gap, like a bad production deploy.
The first run will break. A schema check will fail or a call will hang, and Artificial Analysis measured 56.51 seconds to Neon's first answer token at xhigh, its top reasoning setting. Set generous timeouts, log every failure as a row, and rerun the sheet the day the next model ships.
Practise it step by step: Design an AI-Assisted Workflow
The bottom line
The model you pick this month is temporary. The measurement habit lasts.
Neon shows a mid-tier model can outscore a flagship on an index while costing less per finished task. That lead is two points, and Google alone is shipping a model every 1.75 weeks. We think the durable edge is a small sheet of your own tasks with cost and pass rate beside each one, rerun after every launch. Build it once, keep it current, and let it choose your default.
