Gemini 4 Neon 何時比 Carbon 便宜?解析 2026 年 12 月 1 日調價
Gemini 4 Neon Price: Cheaper Than Carbon After Dec 1?
Gemini 4 Neon 從 2026 年 12 月 1 日起比 Carbon 每 token 便宜 25%,因 Carbon 輸入/輸出價格升至 $4/$20,Neon 維持 $3/$15。
文章將 12 月 1 日前後的 Neon 與 Carbon 價格、快取比例門檻和 API 遷移要求放入具體用量情境,方便團隊核算費用並檢查切換設定。
I spent this week pricing out Google's new Gemini 4 Neon for the kind of agent workloads I run, and the honest answer to "is it cheaper?" is: not today, yes from December 1. That flip catches a lot of teams off guard, so here is the math, the API changes that break old code, and when I would actually pick Neon.
I wrote this with help from AI tools and checked every price and claim against the linked sources myself. The cover image is AI-generated.
Watch the video version: Gemini 4 Neon: Google's Price, What Changed, and When to Use It on my channel Daily AI News By Ali.
Gemini 4 Neon price in brief
Short answer: Gemini 4 Neon costs $3 per 1M input tokens and $15 per 1M output tokens, with cached input at $0.15. Until November 30, 2026 that is 50% more than Gemini 4 Carbon ($2/$10). On December 1, Carbon and Argon double to $4/$20, so Neon becomes 25% cheaper per token. Rerun your cost math before then.
1. The launch: what Gemini 4 Neon is and where it runs
What happened. Google introduced Gemini 4 Neon on November 18, 2026. It has a 1M-token context window and is rolling out in the Gemini API (AI Studio and Android Studio), Gemini for business, Antigravity for paid AI Pro and Ultra users, and the Gemini app for Pro and Ultra. 9to5Google reports that the app picker now reads Auto, 4 Hydrogen, 4 Neon and 4 Carbon, with Neon replacing Argon in the app.
Google's own numbers (company claims, not independent tests): #1 on the Artificial Analysis Coding Agent Index at 58.6, 65.2% on Terminal-Bench 4.1, and 32% lower cost per task than Carbon at the high thinking level. VentureBeat notes that the launch post gives no absolute completion times.
Why it matters. Neon sits between the fast, cheap Hydrogen and the top-end Carbon. For coding agents that run many steps, it is pitched as Carbon-level quality at a lower cost per finished task.
What to watch. Independent benchmarks over the next few weeks. "Cost per task" depends on how many tokens a model burns, and that is exactly the number the launch post does not give.
2. The December 1 price flip for Carbon and Argon
What happened. The Gemini API pricing page lists Gemini 4 Carbon and Argon at $2.00 input, $10.00 output and $0.10 cached "through November 30, 2026", then $4.00, $20.00 and $0.20 "starting December 1, 2026". Neon stays flat at $3/$15/$0.15. Hydrogen on the paid tier is $1.50/$7.50/$0.075.
Here is one monthly workload, 10M input and 2M output tokens, with no caching, at official prices:
| Model | Input $/1M | Output $/1M | Monthly bill (10M in, 2M out) |
|---|---|---|---|
| Gemini 4 Hydrogen | $1.50 | $7.50 | $30 |
| Gemini 4 Carbon (until Nov 30) | $2.00 | $10.00 | $40 |
| Gemini 4 Neon | $3.00 | $15.00 | $60 |
| Gemini 4 Carbon (from Dec 1) | $4.00 | $20.00 | $80 |
Why it matters. If you are on Carbon or Argon and do nothing, the same traffic costs twice as much in ten days. For a lot of SaaS products that is a margin problem, not a rounding error.
What to do. Pull last month's token counts per model from your billing export and multiply by the December rates now, not on December 2.
3. Implicit caching: my break-even number
Google's What's new in Gemini 4 Neon doc says Neon does implicit caching on prompts of 4,096 tokens or more, at a 95% discount, and does not support explicit caching.
The video showed one case: with 8M of the 10M input tokens cached, Neon's bill drops from $60 to about $37. Here is the extra number I worked out for this article. The Neon bill for that workload is:
cost = 2M x $15 (output) + 10M x [$3 x (1 - f) + $0.15 x f]
= $60 - $28.50 x f (f = share of input tokens cached)
Set that equal to Carbon's current uncached $40 and you get f = 0.70. So if at least 70% of your input is a repeated prefix (a long system prompt, tool schemas, a shared codebase context), Neon is already cheaper than today's uncached Carbon bill for this workload. Agent loops that resend the same large context on every step often land there. One-shot chat with short prompts does not, because prompts under 4,096 tokens are not cached at all.
What to do. Put the stable part of your prompt first and the changing part last, so the cached prefix is as long as possible.
4. What changed in the API (the migration traps)
What happened. The same Google doc lists several breaking changes for code that worked on Gemini 3 or other Gemini 4 models:
- Model ID is gemini-4-neon, and it is paid tier only.
- Thinking levels are low, medium, high (the default) and xhigh. Minimal is rejected. Carbon and Argon default to medium and cap at high.
- Sampling knobs are gone. In-range temperature, top_p and top_k are ignored; out-of-range values return a 400.
- Output limits. One shared 1,048,576-token window, a 65,536-token synchronous output cap, and background mode for longer output. Google suggests raising low max_output_tokens caps (for example 8,192) to 65,536 and using timeouts of at least 15 minutes.
- Explicit caching is not supported, so remove cached_content.
- Tools should be declared natively (otherwise you can hit MALFORMED_FUNCTION_CALL), and thought signatures should be passed back unchanged.
Why it matters. The default thinking level is high, so a straight model-ID swap can quietly spend more output tokens than you expect.
What to do. Here is the before/after checklist I use for a config like this (field names follow the doc's wording; check your SDK's exact spelling):
- model: gemini-3.x / gemini-4-carbon + model: gemini-4-neon
- temperature: 0.2 (delete)
- top_p / top_k (delete)
- cached_content: ... (delete, rely on implicit caching)
- max_output_tokens: 8192 + max_output_tokens: 65536
+ thinking_level: test low, medium, high and xhigh; keep the lowest that passes
+ request timeout: >= 15 minutes
Google's own advice is to test all levels and pick the lowest that meets your quality bar. I would run that on 20 to 50 real tasks from your logs, not a toy prompt.
5. How Neon compares with rival models
What happened. VentureBeat lists Claude Sonnet 5.5 and GPT-6.2 Sol (standard, up to 272K) at $2/$10, and Claude Opus 5.5 at $4/$20. Neon's headline rates are 50% above Sonnet and Sol and 25% below Opus. VentureBeat's break-even: Neon needs about one-third fewer tokens than Sol or Sonnet to cost the same. It also quotes Glean saying Neon matched or exceeded Carbon at 3x the speed and 60% of the cost.
Why it matters. Per-token price alone does not decide this. If Neon finishes a coding task in fewer steps, it can still win; if not, Sonnet or Sol stay cheaper.
When to use what (Google's guidance): Hydrogen as the default, Neon for complex high-frequency work, Carbon for the hardest problems, and 3.5 Flash-Lite for high-volume simple calls.
What to watch next
- [ ] Export last month's Gemini token usage and price it at the December 1 rates
- [ ] Measure what share of your input is a repeated prefix over 4,096 tokens
- [ ] Remove temperature, top_p, top_k and cached_content before switching to Neon
- [ ] Test every thinking level on real tasks and keep the cheapest that passes
- [ ] Watch for independent coding-agent benchmarks of Neon vs Sonnet 5.5 and GPT-6.2 Sol
If you want more context on recent AI news for builders, I covered the California lawsuit against OpenAI and Windows coding-agent telemetry and Qwen4-Max, Hy4 Flash and Copilot security earlier on DEV.
More short versions: Gemini Carbon's Dec 1 price change in 60 seconds and why Neon ignores your temperature setting.
FAQ
How much does Gemini 4 Neon cost?
$3 per 1M input tokens, $15 per 1M output tokens and $0.15 per 1M cached input tokens on the paid Gemini API tier, according to Google's pricing page.
Is Gemini 4 Neon cheaper than Gemini 4 Carbon?
Not until November 30, 2026, when Neon is 50% more per token. From December 1, Carbon moves to $4/$20 and Neon is 25% cheaper per token.
Does Gemini 4 Neon support temperature?
No. Google's docs say in-range temperature, top_p and top_k are ignored and out-of-range values return a 400 error. Use thinking levels instead.
Does Gemini 4 Neon support context caching?
Only implicit caching on prompts of 4,096 tokens or more, at a 95% discount. Explicit caching (cached_content) is not supported.
Where can I watch daily AI news?
On my YouTube channel Daily AI News By Ali.
Follow & Connect
- X: @AliMughal887
- GitHub: Muhammad Ali Mughal (@ali111887)
Watch on YouTube: Daily AI News By Ali · Resume Boost AI · PixelValut
來源:DEV Community · dev.to