Skip links

Gemini 4 Neon tested: pricing, benchmarks and Neon vs Carbon

Gemini 4 Neon in white type with an orange glow on navy, above orange filaments fanning out from the bottom edge
Google released Gemini 4 Neon on 18 November 2026, and we tested it two days later.

The answer in brief

Gemini 4 Neon is now the Gemini 4 model to reach for when Hydrogen is not good enough. At the same high setting it ties Carbon on the Artificial Analysis Intelligence Index, finishes in under 40% of the time and, from 1 December 2026, costs 25% less per token. It is not the cheapest route to that score: GPT-6.2 Astra matches it for less.

  • Release: Google released Neon on 18 November 2026, eight days after Carbon and six days after Hydrogen.
  • Access: Google AI Pro (£18.99 a month) and Ultra in the Gemini app and Google Antigravity, the paid Gemini API, and Gemini for business. There is no free tier.
  • Price: $3 for input and $15 for output per million tokens, at US list prices. That is 50% more than Carbon until 30 November 2026 and 25% less from 1 December.
  • Independent score: 61 on the Artificial Analysis Intelligence Index v5.0.1 at its top thinking level, third behind Claude Fable 5.2 (65) and GPT-6.2 Astra (62).
  • Cost per task: at high, Neon, Carbon and GPT-6.2 Astra all score 60 on that index. A task costs $3.74 on Neon, $7.76 on Carbon at its December price and $2.07 on Astra.
  • Our tests: Neon passed all 32 automated checks on our 16 marketing jobs. So did GPT-6.2 Sol, for 27% to 42% of the cost.
  • Invented claims: Neon promised UK delivery unprompted in 3 or 4 of 10 copywriting answers, depending on the thinking level. Sol did so in none.
  • Thinking level: high is Google’s default. On Artificial Analysis’s index, xhigh adds less than two points for 72% more cost. OpenRouter defaults to medium.
  • Made-up answers: on Artificial Analysis’s AA-Omniscience v1.1 test, 14% of Neon’s answers that were not fully correct were wrong, at high or xhigh; the rest were partial or declined. Claude Fable 5.2, Claude Opus 5.5, GPT-6.2 Astra and GPT-6.2 Sol at their top settings scored 50% to 59%.
  • Google’s claims: Neon has the highest reported score on 16 of the 19 rows in Google’s own table. The table leaves out the two models that outrank it, and the headline coding result depends on Google’s own agent software.
  • Argon: Neon has replaced Gemini 4 Argon in the Gemini app. Argon is still in the API, where its price doubles on 1 December 2026.

What is Gemini 4 Neon, and when was it released?

Gemini 4 Neon is the middle model of Google’s Gemini 4 family, released on 18 November 2026 with the ID gemini-4-neon. It sits between Gemini 4 Hydrogen, the fast everyday model, and Gemini 4 Carbon, the largest. Google’s launch post aims it at complex coding, analysis and enterprise workflows, and sums up the position in five words: “Neon balances depth and efficiency”.

Google DeepMind’s model card says Neon is “adapted from Gemini 4 Carbon”. It is the fourth Gemini 4 model in seven weeks, counting from Argon’s announcement on 30 September; the others are Argon, Carbon and Hydrogen.

The specification, from Google’s release notes and Google Cloud model page:

  • Input: text, images, audio, video, PDFs and Office files.
  • Output: text only.
  • Context window: 1,048,576 tokens, shared between what you send, the model’s thinking and its answer. A single response segment can be up to 65,536 output tokens.
  • Thinking levels: low, medium, high (the default) and xhigh.
  • Knowledge cutoff: March 2026.

Can you use Gemini 4 Neon in the UK?

Yes. Google’s UK plans page lists Neon under Google AI Pro at £18.99 a month, and the UK is on Google’s list of supported regions for the Gemini API. The free Gemini app and Google AI Plus do not include it.

Table 1. Where you can use Gemini 4 Neon and what you need. Source: Google’s launch post, UK plans page, API pricing page and Google Cloud model page, and OpenRouter’s listing and FAQ.

If the table extends beyond the screen, scroll sideways to view all columns.

Route Who can use it What it costs
Gemini app Who can use itGoogle AI Pro and Ultra subscribers What it costs£18.99 a month for Pro; Ultra from £79.99 a month
Google Antigravity Who can use itPaid Google AI Pro and Ultra subscribers; businesses through Gemini for business What it costsPro or Ultra for individuals; Gemini for business pricing for organisations
Gemini API and Google AI Studio Who can use itPaid API tier; Pro and Ultra subscribers can also try it in AI Studio What it costs$3 input and $15 output per million tokens
Google Cloud Who can use itGoogle Cloud customers What it costsGoogle Cloud’s published rates
OpenRouter Who can use itAnyone with credits What it costsGoogle’s list price; OpenRouter charges 5.5% when you buy credits

Four details matter for UK teams.

The rollout is gradual. The US version of Google’s Gemini models page lists Hydrogen, Neon and Carbon. The UK English version still lists Argon and the older Pro model. Google’s launch post says Neon is rolling out, so it may not appear in every account at once.

A model you pick by hand starts at a low thinking level. The same help page says a manually selected model defaults to Low. Set High for hard questions and expect it to use more of your allowance. Google’s limits page gives no prompt count for Neon.

There is no UK processing region. Google Cloud offers Neon in three locations: global, US and EU. Argon ran in Google’s London region, europe-west2, and Neon does not: the migration guide tells teams moving from there to switch to one of the three. If a contract requires UK-only processing, raise it with Google before a pilot.

Paid API data is not used for training, but it is logged. Google’s API terms say prompts, files and responses on the paid service are not used to improve its products. They also say Google keeps prompts and responses for a limited period to detect misuse and meet legal requirements. On Google Cloud, the developer guide classes Neon as a Frontier Model under Section 34 of the Service Specific Terms, which covers abuse monitoring. Read the terms for the product you use before sending client material.

How much does Gemini 4 Neon cost?

Neon’s standard Gemini API price is $3 per million input tokens and $15 per million output tokens, with no surcharge for long prompts. These are US list prices in dollars, and Google’s pricing page does not say whether they include tax. Google lists no introductory rate and no free tier for it.

Table 2. API prices in US dollars per million tokens, standard tier. Output includes thinking tokens. Source: Google, Anthropic and OpenAI pricing pages.

If the table extends beyond the screen, scroll sideways to view all columns.

Model Input Output Notes
Gemini 4 Neon Input$3.00 Output$15.00 NotesCached input $0.15; no free tier
Gemini 4 Carbon, to 30 November 2026 Input$2.00 Output$10.00 NotesCached input $0.10
Gemini 4 Carbon, from 1 December 2026 Input$4.00 Output$20.00 NotesCached input $0.20
Gemini 4 Hydrogen Input$1.50 Output$7.50 NotesHas a free tier
Claude Fable 5.2 Input$8.00 Output$40.00 NotesHighest-priced Claude model
Claude Opus 5.5 Input$4.00 Output$20.00 NotesSame as Carbon from December
Claude Sonnet 5.5 Input$2.00 Output$10.00 NotesLower-priced Claude model
GPT-6.2 Astra Input$8.00 Output$40.00 NotesHigher-priced GPT-6.2 model
GPT-6.2 Sol Input$2.00 Output$10.00 NotesLower-priced GPT-6.2 model

Batch and Flex requests are half price, Priority is $5.40 and $27, and cached input is $0.15. Caching is automatic only, on repeated openings of at least 4,096 tokens. Grounding with Google Search includes 5,000 free search queries a month, shared across Gemini 3 and newer models, then costs $14 per 1,000 queries, up to 1,500 a day. One request can run several queries.

As a worked example, a request with 100,000 input tokens and 20,000 output tokens costs $0.60 on Neon. The same token counts cost $0.30 on Hydrogen, $0.40 on Carbon until 30 November and $0.80 from 1 December. But models do not use the same number of tokens for the same job, and that is where the bill is decided.

Why the token price is not the bill

Thinking tokens are billed as output, and Neon thinks a lot. On the Artificial Analysis test suite, Neon at its default level wrote about 90,000 output tokens per task, and just over half were reasoning the user never sees. A lower token price does not mean a lower bill.

Does Neon cost 32% less than Carbon?

From 1 December, more than that. Until then, barely. Google’s launch post says Neon at high has a 32% lower cost per task than Carbon, next to its Artificial Analysis Intelligence Index score, but does not say which Carbon setting or price it compared with. Artificial Analysis measured $3.74 a task for Neon at high and $7.76 for Carbon at high, which is 52% less, but it costs Carbon at the December price. At Carbon’s price until 30 November that halves to about $3.88 by our arithmetic, within a few per cent of Neon.

Artificial Analysis’s launch analysis adds that Neon at xhigh uses about 15% more output tokens per task than Carbon at high, so at those settings the saving comes from the lower token price alone. At high the picture is different: on Artificial Analysis’s data, Neon writes about 90,000 output tokens per task against about 123,000 for Carbon, so it saves on both counts.

What benchmarks has Google published for Gemini 4 Neon?

Google’s model card compares Neon with Claude Opus 5.5 and GPT-6.2 Sol on 19 rows, and Neon has the highest reported score on 16. It loses two coding tests narrowly to Opus 5.5 and a chart-reading test to Sol, and three rows have no score for one of the rivals.

Table 3. Selected rows from Google’s benchmark table for Gemini 4 Neon. Source: Google DeepMind’s model card.

If the table extends beyond the screen, scroll sideways to view all columns.

Benchmark What it tests Neon Claude Opus 5.5 GPT-6.2 Sol
AutomationBench What it testsBusiness automation across apps Neon56.3% Claude Opus 5.542.5% GPT-6.2 Sol41.3%
GraphWalks, 256k to 1M tokens What it testsReasoning over very long inputs Neon95.2% Claude Opus 5.566.8% GPT-6.2 Sol80.1%
Chartography What it testsReading dense charts Neon73.5% Claude Opus 5.566.3% GPT-6.2 Sol75.1%
Terminal-bench 4.1 What it testsCoding in a terminal Neon65.2% Claude Opus 5.565.8% GPT-6.2 Sol61.5%
SWE-sweep, each vendor’s own agent software What it testsAgentic coding Neon39.0% Claude Opus 5.510.1% GPT-6.2 Sol8.3%
SWE-sweep, the same agent software for all What it testsAgentic coding Neon12.8% Claude Opus 5.53.6% GPT-6.2 Sol11.6%

Read the table with two cautions.

The comparison leaves out the models that beat Neon. Claude Fable 5.2 and GPT-6.2 Astra, the two models above Neon on the Artificial Analysis index, are not in Google’s table.

The biggest coding lead comes from Google’s own agent software. On SWE-sweep, Neon scores 39.0% inside Google Antigravity and 12.8% with the neutral setup all three models share, where it leads GPT-6.2 Sol by 1.2 points. Google’s figures do not separate the model’s share from the software’s, but the gap is a reason not to expect the higher figure in another coding tool.

The people who run the tests confirm some of Google’s claims, with one caveat. Vals AI ranks Neon, tested at xhigh, second of 42 models on its Vals Index at 70.87%, and first on its legal research and tax tests. Second place is thin, though. Vals’s leaderboard has Claude Fable 5.2 on 71.31% and Claude Sonnet 5.5 on 70.85%, and Vals puts the margin of error on Neon’s score at about one point. Both gaps are smaller than that margin.

In Artificial Analysis’s launch analysis, Neon at xhigh in Antigravity scores 59 on its Coding Agent Index, in line with Claude Code running Fable 5.2 at xhigh or Opus 5.5 at max (58 each), for $6.86 a task against $15.47 and $12.61. One claim cannot be checked: Google’s launch post reports its lowest fabrication rate yet in internal Antigravity testing and publishes no figure.

How does Gemini 4 Neon compare with Claude Fable 5.2, Claude Opus 5.5 and GPT-6.2?

On the Artificial Analysis Intelligence Index, Neon ranks third at its top setting, behind Claude Fable 5.2 and GPT-6.2 Astra. It beats the Claude models on cost for the score and loses to OpenAI’s.

Table 4. Selected models on the Artificial Analysis Intelligence Index v5.0.1. Carbon is costed at $4 and $20 per million tokens, its price from 1 December 2026. Costs and times are averages over long benchmark tasks, not short prompts, and they change as the models are re-measured. Source: Artificial Analysis.

If the table extends beyond the screen, scroll sideways to view all columns.

Model and setting Index Cost per task Minutes per task
Claude Fable 5.2 (max) Index65 Cost per task$26.54 Minutes per task42.0
GPT-6.2 Astra (max) Index62 Cost per task$4.00 Minutes per task14.2
Gemini 4 Neon (xhigh) Index61 Cost per task$6.44 Minutes per task21.5
Gemini 4 Neon (high) Index60 Cost per task$3.74 Minutes per task14.8
Gemini 4 Carbon (high) Index60 Cost per task$7.76 Minutes per task39.9
GPT-6.2 Astra (high) Index60 Cost per task$2.07 Minutes per task6.3
Claude Opus 5.5 (max) Index59 Cost per task$10.04 Minutes per task25.7
GPT-6.2 Sol (max) Index57 Cost per task$1.01 Minutes per task9.6
Claude Sonnet 5.5 (max) Index55 Cost per task$10.02 Minutes per task32.9

Against Claude, Neon is good value. At its default level it scores 60 against 59 for Opus 5.5 at max, for a little over a third of the cost.

Against GPT-6.2, it is not. Astra at high matches Neon at high for $2.07 a task against $3.74, in six minutes against 15. Sol at max scores 57 for $1.01. Our GPT-6.2 Astra and Sol comparison covers those two in detail.

Where Neon separates from both is caution.

Table 5. Knowledge and made-up answers on AA-Omniscience v1.1. The made-up answer rate is the hallucination rate: of the responses that were not fully correct, the share that were wrong, as opposed to partial or not attempted. Source: Artificial Analysis’s Neon page and AA-Omniscience page.

If the table extends beyond the screen, scroll sideways to view all columns.

Model and setting Answered correctly Made-up answer rate
Gemini 4 Carbon (medium) Answered correctly48.9% Made-up answer rate4.2%
Gemini 4 Carbon (high) Answered correctly58.3% Made-up answer rate11.4%
Gemini 4 Neon (high) Answered correctly59.4% Made-up answer rate13.9%
Gemini 4 Neon (xhigh) Answered correctly60.5% Made-up answer rate14.4%
Claude Opus 5.5 (max) Answered correctly65.5% Made-up answer rate50.3%
GPT-6.2 Sol (max) Answered correctly61.6% Made-up answer rate55.4%
GPT-6.2 Astra (max) Answered correctly68.6% Made-up answer rate59.1%
Claude Fable 5.2 (max) Answered correctly79.1% Made-up answer rate59.3%

When Neon cannot give a fully correct answer, it mostly gives a partial one or none. The Claude and GPT models at their top settings guess wrong more than half the time in the same position. The trade-off is knowledge: Fable 5.2 answers 79% of the questions correctly and Neon 60%. For fact-heavy marketing work, a model that declines is easier to check than one that guesses. Even so, about one in seven of its answers that are not fully correct is wrong, so the checking does not go away.

A separate test found a split between factual and sales copy. Spring Prompt gave 35 models supplier data with some fields blanked out, plus product photos, and counted a listing only if it was ready to publish in all three runs. It ran Neon at OpenRouter’s default, medium. Neon came first on the factual brief, with 44 of 56 listings passing. On a sales brief it was 13th, with 32.1% passing against 75% for GPT-6.2 Sol. Spring Prompt sells model testing, and this is one vendor’s benchmark. We saw a related weakness in our own Carbon and Hydrogen tests, where both models added delivery, origin or manufacturing claims the brief never made. Check every product promise in sales copy.

How did Gemini 4 Neon do in our own tests?

Neon passed every automated check in our tests, and so did GPT-6.2 Sol, for 27% to 42% of Neon’s cost. On 20 November 2026 we ran Neon at medium and high, and GPT-6.2 Sol at max, through OpenRouter on the 16 marketing jobs we used for Carbon and Hydrogen, two passes each. The jobs cover copywriting, reporting, ad analysis, translation, product data, a 99,000-token document, false premises and instructions hidden in the input.

Table 6. AIWIZ’s 16-job test through OpenRouter, two passes per setting. Costs are what OpenRouter billed; Carbon’s December figures are our estimates at its new price. GPT-6.2 Sol was sent the max setting. Carbon was tested on 10 November and Hydrogen on 12 November 2026, on the same jobs; for those two, checks passed counts answers confirmed by reading where an automatic check failed.

If the table extends beyond the screen, scroll sideways to view all columns.

Model and setting Automated checks passed Cost of all 32 answers Median seconds, short jobs Copy with an unprompted delivery or origin claim
Gemini 4 Neon, high Automated checks passed32/32 Cost of all 32 answers$2.36 Median seconds, short jobs20.2 Copy with an unprompted delivery or origin claim4/10
Gemini 4 Neon, medium Automated checks passed32/32 Cost of all 32 answers$1.54 Median seconds, short jobs17.0 Copy with an unprompted delivery or origin claim3/10
GPT-6.2 Sol, max Automated checks passed32/32 Cost of all 32 answers$0.64 Median seconds, short jobs16.8 Copy with an unprompted delivery or origin claim0/10
Gemini 4 Carbon, high Automated checks passed31/32 Cost of all 32 answers$2.79, or $5.57 at its December price Median seconds, short jobs56.5 Copy with an unprompted delivery or origin claim4/10
Gemini 4 Carbon, medium Automated checks passed32/32 Cost of all 32 answers$1.12, or $2.24 at its December price Median seconds, short jobs20.3 Copy with an unprompted delivery or origin claim4/10
Gemini 4 Hydrogen, high Automated checks passed32/32 Cost of all 32 answers$0.84 Median seconds, short jobs9.6 Copy with an unprompted delivery or origin claim3/10

Neon is the better Carbon. At high it finished our long-document job both times, where Carbon at high once stopped at its output limit with no answer. Across all 32 calls it took about a third of Carbon’s time and, at Carbon’s December price, 58% less money.

It is not better than Sol. Both passed every automated check. Neon cost two to four times as much, partly because its tokens cost 50% more and mostly because it thought for longer: it wrote an average of 2,765 output tokens per short job at high, against 1,002 for Sol, and nearly all of the difference was thinking.

It still adds claims nobody asked for. None of our copywriting prompts mention delivery or where the lamps are made. Neon promised UK delivery in meta descriptions and ad headlines in 3 of 10 copywriting answers at medium and 4 of 10 at high, and called the lamps handmade or handcrafted in seven of its 20 copywriting answers. Sol did neither. Carbon and Hydrogen made the same kind of additions in our earlier tests, so check every promise in Gemini copy before it goes live.

Gemini 4 Neon vs Carbon: which should you use?

Use Neon unless you have a task where Carbon measurably does better. At their default settings Neon scores slightly higher on the independent index, takes half the time and, from 1 December 2026, costs 43% less per task.

Table 7. Gemini 4 Neon and Carbon on the Artificial Analysis Intelligence Index v5.0.1 and its component tests. Carbon is costed at its price from 1 December 2026. Source: the Artificial Analysis page for Neon.

If the table extends beyond the screen, scroll sideways to view all columns.

Measure Neon at high (default) Carbon at medium (default) Carbon at high
Intelligence Index Neon at high (default)60 Carbon at medium (default)58 Carbon at high60
Cost per index task Neon at high (default)$3.74 Carbon at medium (default)$6.59 Carbon at high$7.76
Minutes per index task Neon at high (default)14.8 Carbon at medium (default)28.8 Carbon at high39.9
Time to first answer token Neon at high (default)41 seconds Carbon at medium (default)33 seconds Carbon at high75 seconds
Terminal-Bench 4.0 Neon at high (default)61.6% Carbon at medium (default)60.6% Carbon at high65.2%
GDP.pdf, reasoning over documents Neon at high (default)33.4% Carbon at medium (default)29.6% Carbon at high28.4%
Made-up answer rate, AA-Omniscience v1.1 Neon at high (default)13.9% Carbon at medium (default)4.2% Carbon at high11.4%

Carbon keeps two advantages. At high it is about four points ahead on Terminal-Bench 4.0, the terminal coding test. At medium it almost never guesses: its made-up answer rate is 4.2% against Neon’s 13.9%. Until 30 November 2026 Carbon is also cheaper per token: at that price, by our arithmetic, a task would cost about $3.29 at medium and $3.88 at high, against $3.74 for Neon at high.

Neon is faster by a wide margin and has a fourth thinking level. It also comes with the £18.99 Google AI Pro plan, where Carbon needs Ultra at £79.99.

If you already use Carbon through the API, our Carbon tests found medium was its best-value level. Run the same jobs on Neon at high before 1 December and compare the bills.

Gemini 4 Neon low vs medium vs high vs xhigh: which thinking level should you use?

Start at high, the API default, and move down to medium for interactive work. On Artificial Analysis’s index each step up costs 1.7 to 2.5 times as much per task, and the last step adds the fewest points.

Table 8. Gemini 4 Neon at each thinking level on the Artificial Analysis Intelligence Index v5.0.1 and three of its component tests. Source: Artificial Analysis.

If the table extends beyond the screen, scroll sideways to view all columns.

Thinking level Index Cost per task Minutes per task Terminal-Bench 4.0 GDP.pdf Made-up answer rate
Low Index44 Cost per task$0.75 Minutes per task3.7 Terminal-Bench 4.032.8% GDP.pdf31.4% Made-up answer rate18.8%
Medium Index55 Cost per task$1.90 Minutes per task8.3 Terminal-Bench 4.054.5% GDP.pdf32.8% Made-up answer rate16.2%
High (default) Index60 Cost per task$3.74 Minutes per task14.8 Terminal-Bench 4.061.6% GDP.pdf33.4% Made-up answer rate13.9%
Xhigh Index61 Cost per task$6.44 Minutes per task21.5 Terminal-Bench 4.063.6% GDP.pdf28.0% Made-up answer rate14.4%
  • Low is what Google’s guide suggests for fast chat and high-throughput processing, and Google says it is still capable on agentic tasks. Its terminal coding score is almost half its score at high, so test it before you run coding agents on it.
  • Medium suits people waiting at a screen. It costs half as much as high and takes about half the time, for five fewer index points. In our own 16-job test it passed every automated check for 35% less than high.
  • High is where Neon matches Carbon at high on the overall index.
  • Xhigh is rarely worth it. It adds less than two points for 72% more cost, and it scores lower on the document test than any other level.

A second source points the same way on xhigh and is kinder to low. In Zapier’s AutomationBench report, which runs 657 business automation tasks at each level, Neon passed 49.6% at low, 53.3% at medium, 54.3% at high and 56.3% at xhigh. Zapier’s cost estimates were $0.40, $0.58, $0.88 and $1.69 a task. The top level buys two more points for nearly double the cost of high, and low gives up less on automation than it does on coding.

If you move code from Argon, set the level yourself. Argon defaulted to medium and Neon defaults to high, so a straight swap of the model ID can raise the bill. OpenRouter works the other way: a request that sets no level gets medium, not Google’s high.

What happens to Gemini 4 Argon and Gemini 3.1 Pro?

Neon replaces both in the Gemini app, and Google’s migration guide points Argon users to it. Google’s US help page now lists Hydrogen, Neon and Carbon alongside Auto, and 9to5Google reported on launch day that 3.1 Pro is being removed.

In the API nothing has been switched off. Google’s deprecations page shows no shutdown date for either model, and Google Cloud’s lifecycle page keeps Argon until 13 October 2027 or later. Neon has no such guarantee: it is on Google Cloud’s shorter-availability list, with at least 45 days’ notice before any retirement.

The reason to move is price. Argon doubles to $4 and $20 per million tokens on 1 December 2026, when it starts to cost more per token than Neon. It also scores lower: 50 at high on the Artificial Analysis Intelligence Index, against 60 for Neon at high.

Google’s migration guide lists changes that break or cost money if missed:

  1. Change the model ID to gemini-4-neon and the location to global, US or EU.
  2. Set the thinking level yourself, and replace thinking_budget and the MINIMAL level, which now return an HTTP 400 error.
  3. Remove explicit caching. Neon only caches automatically.
  4. Stop relying on temperature, top_p and top_k. Neon ignores custom values and returns an error for out-of-range ones.

One thing to test first: Argon at high gave a wrong answer on 4.3% of the knowledge questions it did not get fully right, and Neon at high on 13.9%. If you chose Argon because it rarely guesses, check Neon on your own material.

What are the known limits of Gemini 4 Neon?

Neon’s main limits are a slow start, heavy token use and the lack of a free tier. In practice:

  • It is slow to start above medium. Artificial Analysis measured about 41 seconds to the first answer token at high and 57 at xhigh, against 20 at medium. That matters more for live chat than for batch work.
  • It writes a lot. Across the Artificial Analysis test run, Neon at xhigh produced 283 million output tokens, against a median of 65 million for comparable models. You pay for all of them.
  • It still makes things up, and its knowledge stops at March 2026. For anything recent, give it a current source in the prompt or turn on a tool such as Grounding with Google Search, which is billed per query after the free allowance.
  • The safety report is still to come. Google has published the model card and says the full report will follow.

How should you test Gemini 4 Neon on agency work?

Run it on work you can mark, against the model you use now. Below is one exercise to copy and two briefs to adapt. We ran the exercise; the two briefs are tests to run, not results.

A paid media reporting exercise with an answer key

A model can write a fluent report and still recommend the wrong channel. This exercise checks whether it keeps ad platform conversions separate from unique qualified leads in the CRM. The data is made up. Give both tables and the prompt to Neon at high, Neon at medium and your current model, each in a fresh session with the same tools.

Table 9. The September platform export for the exercise.

If the table extends beyond the screen, scroll sideways to view all columns.

Channel Spend Platform conversions
Google Ads Spend£1,100 Platform conversions26
Paid social Spend£600 Platform conversions12
Table 10. The CRM export for the exercise.

If the table extends beyond the screen, scroll sideways to view all columns.

Row Lead ID Created on Source Stage
R1 Lead IDL01 Created on3 Sep SourceGoogle Ads StageQualified
R2 Lead IDL02 Created on8 Sep SourceGoogle Ads StageBooked
R3 Lead IDL02 Created on9 Sep SourcePaid social StageBooked
R4 Lead IDL03 Created on15 Sep SourceGoogle Ads StageSpam
R5 Lead IDL04 Created on1 Oct SourceGoogle Ads StageQualified
R6 Lead IDL05 Created on18 Sep SourcePaid social StageQualified
R7 Lead IDL06 Created on25 Sep SourcePaid social StageBooked
R8 Lead IDL07 Created on27 Sep SourcePaid social StageNew
R9 Lead IDL08 Created on29 Sep SourceUnknown StageQualified

The prompt to send with both tables:

Review September only. Keep the earliest September CRM record for each lead ID across all sources, then count only Qualified and Booked as qualified leads. Report unknown-source qualified leads separately. For each paid channel, show spend per platform conversion and spend per unique qualified CRM lead, rounded to two decimal places. Cite the CRM row IDs behind each lead count. In two sentences, explain whether the channel ranking changes and what these exports can establish about the gap.

Table 11. The answer key. One more qualified lead, R9, is left out of both channel counts because its source is unknown.

If the table extends beyond the screen, scroll sideways to view all columns.

Channel Spend per platform conversion Qualified CRM leads Spend per qualified CRM lead
Google Ads Spend per platform conversion£42.31 Qualified CRM leads2 (R1, R2) Spend per qualified CRM lead£550.00
Paid social Spend per platform conversion£50.00 Qualified CRM leads2 (R6, R7) Spend per qualified CRM lead£300.00

Google Ads is cheaper per platform conversion and paid social is cheaper per qualified lead, so the ranking reverses. R3 repeats lead L02, R5 falls in October, and R4 and R8 are not qualified. The exports show a gap between the platform and CRM counts. They cannot show what caused it.

Mark each run on four checks:

  1. Both platform costs are right.
  2. Both lead counts and costs are right, with the right rows cited.
  3. The unknown-source lead is kept separate.
  4. The explanation spots the reversal without blaming the gap on a cause the data does not show, such as a tracking fault.

The fourth check tests the same habit as the made-up answer rate, on your own kind of data.

When we ran it on 20 November 2026, three times each on Neon at high and medium, Hydrogen at high and GPT-6.2 Sol at max, all 12 answers got every number and row right and spotted the reversal. The differences were in the explanation. In four of Neon’s six answers and one of Hydrogen’s three, the explanation counted the October lead, R5, as part of the September gap, which the data does not show. A fifth Neon answer mentioned R5 in a way that can be read either way. None of Sol’s did. On average Neon cost $0.040 an answer at high and $0.034 at medium, against $0.010 for Sol, with median times of 23, 21 and 15 seconds.

Two briefs to adapt

  • SEO content maintenance. Give it a published article and the current versions of its sources. Ask for a list of claims that need changing, each tied to the passage that proves it. Have a person check the dates and the exact wording.
  • A web build. Give Neon and your current model the same mobile-first brief for a lead generation page with a quote calculator and a fixed brand kit. Check the layout at 390 px, keyboard use, the calculator’s arithmetic, and that a submitted form fires one conversion event even when someone double-clicks. Run it in the coding tool your team already uses, because Neon’s coding scores change with the software around it.

For every job, record four things on each model: whether the first answer was right, how long it took, what the API charged and how many minutes a person spent correcting it. Give the model read-only access to source material, and have a person approve anything before it is sent or published.

Should you switch to Gemini 4 Neon?

It depends on what you use now.

  • You pay for Google AI Pro. There is little to decide: on Artificial Analysis’s index, Neon scores well above Hydrogen, the other model on your plan. Use it for analysis and long documents, raise the thinking level to High for hard questions, and leave everyday prompts on Auto or Hydrogen.
  • You call Argon through the API. Move before 1 December 2026, when Argon’s price doubles, and set the thinking level yourself.
  • You call Carbon through the API. Run your own tasks on Neon at high. On Artificial Analysis’s benchmark at high, the cost per task is about the same until 30 November and Neon is much faster; from 1 December Neon costs about half.
  • You use Hydrogen for volume. Stay there. Hydrogen at high costs half as much per task as Neon at high on Artificial Analysis’s benchmark. Send Neon the jobs Hydrogen gets wrong.
  • You use Claude or GPT-6.2. Neon is not the cheapest way to this level of quality: in our tests GPT-6.2 Sol passed the same checks for 27% to 42% of the cost. Neon’s case is narrower: it gives far fewer wrong answers on Artificial Analysis’s knowledge test, it reads a million tokens at one flat price, and it is built into the Gemini app and Antigravity.

Whatever you choose, do not pick on token price. On the same index, a task costs $1.01 on GPT-6.2 Sol at $2 and $10 per million tokens and $10.02 on Claude Sonnet 5.5 at the same list price. Measure cost per finished job on your own work.

Need help choosing or testing an AI workflow?

We test new models on real marketing jobs before recommending them to clients, and we can run the same checks on one of yours. Talk to AIWIZ or read about our AI adoption services.

Explore
Drag