Gemini 4 Neon vs GPT-6.2 and Claude: Price and Speed

Google is naming models after the periodic table now, and Gemini 4 Neon is the newest element. It landed on November 18, the fourth Gemini 4 release in seven weeks. Google pitches it as the middle child: smarter than Hydrogen, cheaper and faster than Carbon. The real question is how it stacks up against the best from OpenAI and Anthropic, so I lined up the prices and the independent scores.
What Gemini 4 Neon is
Neon is a reasoning model with a 1 million token context window and four thinking levels: low, medium, high (the default) and xhigh. It takes text, images, audio and video, and returns text. API pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.15. Google's model card says it was adapted from Carbon.
The odd bit: Neon beats its bigger sibling. Artificial Analysis scores Neon (xhigh) at 61 on its Intelligence Index, one point above Carbon, at about 17% lower cost per task. Carbon's normal price is $4/$20, so the cheaper chemical wins.
The comparison table
Prices come from each vendor's pricing page. Index, cost per task and time per task come from Artificial Analysis, which runs its own tests, at each model's top effort setting.
| Model | Input / 1M | Output / 1M | AA Index | Cost per task | Time per task |
|---|---|---|---|---|---|
| Claude Fable 5.2 (max) | $8 | $40 | 65 | $26.54 | 41.8 min |
| GPT-6.2 Astra (max) | $8 | $40 | 62 | $4.00 | 13.9 min |
| Gemini 4 Neon (xhigh) | $3 | $15 | 61 | $6.44 | 21.5 min |
| Claude Opus 5.5 (max) | $4 | $20 | 59 | $10.04 | 25.2 min |
| GPT-6.2 Sol (max) | $2 | $10 | 57 | $1.01 | 8.7 min |
Look at the Astra row twice. Its tokens cost more than double Neon's, yet a task costs less, $4.00 against $6.44. Astra simply uses fewer tokens to get there. Sticker price per token tells you very little. Cost per finished task is the number that hits your invoice.
Where Neon actually wins
Neon is not the top scorer overall. Fable 5.2 is, and you pay for it. Neon earns its spot in narrower places:
- Business workflows. On Artificial Analysis's AutomationBench run, Neon (xhigh) scores 79.3%, ahead of Astra at 77.2% and Fable at 76.4%.
- Finance and law. Vals AI ranks Neon #2 of 42 on the Vals Index at 70.87%, and #1 on its Legal Index and Finance Index.
- Coding agents. Artificial Analysis gives Neon in Antigravity CLI a 59 on its Coding Agent Index, close to Claude Opus 5.5 (max) and Fable 5.2 (xhigh) at 58, for $6.86 per task against $12.61 and $15.47.
- Long context. Google reports 95.2% on GraphWalks at 256K to 1M tokens, against 66.8% for Opus 5.5. That one is Google's own number, so treat it as a claim until someone else runs it.
A quick example
Say your team runs 1,000 agent tasks a month that look like the ones Artificial Analysis tests. At top effort, that's roughly $1,010 on Sol, $4,000 on Astra, $6,440 on Neon, $10,040 on Opus 5.5 and $26,540 on Fable 5.2. Your tasks are not their tasks, so test on your own work before you move a budget. Still, a 26x spread between the cheapest and priciest option is hard to ignore.
My take
I think Neon is a strong pick if you live in Google's stack, feed it long documents or video, or run agents that grind through business workflows. If your work is mostly coding and you only care about price per finished task, Astra and Sol make a better case on these numbers. Fable 5.2 is the one you call when being right matters more than the bill.
The surprising part to me is not Neon. It's that the model with the highest token price in this table, GPT-6.2 Astra, is one of the cheapest to actually run. So when you compare models, do you look at the price per token or the price per finished task?