TRUEFUTURE MEDIA INSIGHTS · NOVEMBER 19, 2026
Google introduced Gemini 4 Neon on November 18. If your team uses Gemini 4 Carbon to turn product documentation into customer-facing copy, the price comparison changes on December 1. The model that costs less per token won’t necessarily cost less to put through review.
Which model should you test first?
Test Neon at its default High thinking level against your current Carbon setup. On the standard paid Gemini API, Neon costs $3 per million input tokens and $15 per million output tokens. Carbon costs $2 and $10 through November 30, then $4 and $20 from December 1. Artificial Analysis rounded Neon High and Carbon High to the same score, but that doesn’t tell you which model will handle your source material better. Use the same examples and review checklist for both. Count corrections, retries and review time. Keep a higher thinking level for cases where it fixes a repeatable error without creating a new one.
For a complex product, that error might be one line: a draft that says a retrofit kit works on all units when the manual excludes older serial numbers. A buyer could order the wrong part on the strength of that sentence. Start with the evidence you already use for explaining product compatibility, then see which model preserves the exclusion.

What changes on December 1?
The launch announcement put Neon in the Gemini API through AI Studio and in Gemini for business, with access for Pro and Ultra subscribers in the Gemini app and Antigravity. The numbers below are for the API. They aren’t subscription prices or quotes for other service tiers.
| Model and period | Uncached input | Output, including thinking |
|---|---|---|
| Neon, current listed rates | $3 | $15 |
| Carbon through Nov. 30 | $2 | $10 |
| Carbon from Dec. 1 | $4 | $20 |
At equal token counts, Neon’s standard rates are 50% higher than Carbon’s introductory rates and 25% lower than Carbon’s December rates. A hypothetical run with 100,000 uncached input tokens and 10,000 billed output tokens costs $0.45 on Neon, versus $0.30 and then $0.60 on Carbon. This holds the workload constant and leaves out caching, tool fees and retries. In practice, one model may use more billed tokens to finish the same brief. Check Google’s current API rate card before repricing a workflow.
What does higher thinking buy you?
A thinking level controls how much reasoning a model can do before returning an answer. Artificial Analysis reported an Intelligence Index score of 60 for Neon at its default High setting and 61 at Xhigh. On that suite, time per index task rose from 14.8 to 21.5 minutes, about 45%. This is a benchmark trade-off, not a forecast for a product-copy task.
Carbon High also rounds to 60 in that comparison. A shared rounded score can hide different performance on individual tests. Your buyer will never see the benchmark. They will see whether the answer applied to the unit they own, and whether they can verify it.
When to try a higher setting
First fix any missing source or unclear instruction. If the draft still misses a rule that lives across two documents, rerun that case at Xhigh. Keep the setting for that task if the correction survives review and the added time is acceptable.
How do you test both models on product work?
- Start with 20 examples your team already knows how to check. Mix common questions, exceptions and questions the material cannot answer. Use approved or anonymized sources, and keep confidential material out of a test environment that isn’t cleared for it.
- Give both models the same evidence and output rules. Ask for a buyer-ready answer, an exact source for each consequential claim and a clear note when evidence is missing. Record each model and thinking setting with the result.
- Review against the reference, not the writing style. Give reviewers the expected answer and relevant source excerpts. Hide the model label if workable. Log whether the result preserved the deciding condition and how much time the reviewer needed.
Example: a retrofit-kit compatibility answer
A fictional product bulletin says kits work only on units with serial number 40000 or higher and bracket revision B. Put those two conditions in separate sources. Ask for an answer to a buyer with serial number 41204 but no bracket information. A passing draft explains that the serial number meets one condition and asks for the bracket revision before recommending the kit. A confident “yes” fails.
That same buyer question belongs in your buyer education and proof plan. If your team cannot point to the rule in the first place, a new model cannot make it true.
What does an approved draft cost?
Divide total test spend by the number of outputs that pass review. Include the spend for discarded drafts and retries. Track human review minutes next to the API total, or add a documented labor rate if that helps your purchasing decision. Track turnaround from submission to usable output for each case. With a small-run test, report the slowest cases as well as the median so delays don’t disappear into an average.
Worked example: the cheaper run can cost more
In an illustrative pilot, one setup costs $8 and clears 10 of 20 drafts: $0.80 per passing draft. Another costs $10 and clears 16: about $0.63 per passing draft. Review labor is excluded in both. If the second requires substantially longer review, the ranking may change once you count it.
Keep material errors separate from style edits. A missing exclusion, unsupported safety claim or mismatched product version can create a rework or return that your per-draft average won’t capture.
When should you switch?
Shift repeatable tasks to Neon if it meets your claim-accuracy standard and sustains a better cost or turnaround per approved draft. Keep Carbon where it catches an important exception more reliably. Recheck during the first December billing cycle with real usage, since the benchmark comparison cannot tell you the token mix for your sources.
Three checks before you migrate
- Can a reviewer trace each compatibility, performance or safety claim to an approved source?
- Have you counted failed runs and human correction time?
- Does the fallback catch the mistakes that would change a buyer’s decision?
Frequently asked questions
Does Carbon become more expensive on December 1?
Yes. Standard paid input and output rates go from $2/$10 to $4/$20 per million tokens. Other tiers have their own rates.
Does Neon Xhigh cost more per output token?
Under the listed standard rate, the output token price is the same. Thinking tokens are billable, so a longer reasoning run can cost more overall.
Is this a reason to move every workflow to Neon?
No. Re-run your own cases and keep a human checkpoint where a faulty claim would affect safety, product fit or purchase.
Give the model better evidence to work with
A model test can expose a missing answer in your manual, a compatibility note your sales team has to repeat, or a demo that skips the hardest buyer question. Put that evidence in the material buyers see before they ask for a quote.
If your team knows the product well but its content keeps leaving buyers with the same questions, start with one decision you need to make easier.
Request a Fit Review
