Last month, a developer used Gemini 4 Argon to check an old project in Antigravity and posted that it found a bug that was not seen in other cutting-edge models. However, in the same conversation, it also made up several calls that did not exist in the code. Positive and negative reviews are written in the same post.

I have always remembered this post because on November 18th, Google released another Gemini: Gemini 4 Neon. The fourth Gemini 4 model in seven weeks, the main focus is on writing code as an agent, and the comparison object directly named Claude Opus 5.5.

No one can currently give an answer as to whether Neon has inherited Argon's problem of "accurately finding and ruthlessly editing". What can be found are several other accounts.

In Google’s 19-item list, which items are losing?

According to Google’s official table, Neon outperformed Claude Opus 5.5 and GPT-6.2 Sol in 16 out of 19 tests.

If you look closely, there are two discounts. First, there is an opponent in the table who has no points in 3 of the items. The blank spaces are not counted as wins and losses. The 16 wins are counted after deducting the blank spaces from the 19 items. Second, among the three items that Neon lost, two were programming tests, both lost to Claude; the remaining one was lost to OpenAI based on the chart.

A model that focuses on programming intelligent agents lost twice to its opponents on programming questions chosen by the manufacturer.

The arrangement of third parties is also different. In the comprehensive index of Artificial Analysis, Neon scored 61 points, ranking third, followed by GPT-6.2 Astra and Claude Fable 5.2, both of which are larger and slower. Opus 5.5 is not ahead of it, and Google is right about that.

When programming an agent, the bill difference nearly doubles

Where Neon really stands up is in cost. Artificial Analysis ran a set of tasks for programming agents alone. Neon and Opus 5.5 were tied. The cost per task was US$6.86 versus US$12.61. Opus was 84% ​​more expensive.

The price tag itself is only a little over 30% off. Neon’s input is $3 and output is $15 per million tokens; Opus 5.5 is $4 and $20, which was once lower than the previous generation Opus 5’s $5 and $25 when it was released on September 22. The unit price difference is 25%, and the task bill difference is 84%. The extra part comes from Opus spitting out more tokens when thinking.

Opus has a history of this. Its max file outputs an average of about 173,000 tokens per question, and it took 218 million tokens to run the AA complete index. Anthropic later released Fable 5.2, which is even better. The same set of questions failed 275 million, so Opus is still the restrained one in its own camp.

Gemini 4 Neon Claude Opus 5.5
release date 2026-11-18 2026-09-22
List price (per million tokens, input/output) 3 / $15 4 / $20
cache read The official page is not listed separately $0.20
Bulk interface Not found half price
Context/Single Maximum Output About 1 million / About 1 million token 1 million / 128,000 tokens
AA programming agent cost per task $6.86 $12.61

Opus account has two buttons to pull back

If you compare prices directly, you will miss out on Opus’s two discounts.

Cache reads only cost $0.20 per million tokens, which is one twentieth of the normal input price. The programming agent has to re-feed the context of the same warehouse every round. After hitting the cache, the cost on the input side is almost negligible, and the rest of the bill is mainly output. Batch interfaces are directly discounted by 50%. For non-urgent work such as code review and batch replenishment testing at night, after Batch, the unit price of Opus dropped to US$2 and US$10, which is lower than the price of Neon.

There's a button on the opposite side, too: fast mode, which outputs up to 2.5x faster and flips the price to $8 and $40. Drive it in a hurry, and that 84 percent gap becomes more than three times that.

Therefore, the phrase "Neon is half the price" is only true in scenarios where real-time interaction and cache hit rate are not high.

Neon’s shortcoming is not the score

My biggest concern with Neon is pacing.

Argon was released on September 30. At that time, it was only open to a small number of trusted testers and security partners. A review author tried three model IDs and all returned 404. After that, Carbon, Hydrogen, and Neon were released one after another, seven to four models, and the prices were divided into promotional prices and regular prices. The Argon set’s promotional prices of $2 and $10 will not end until November 30. To someone writing a one-off script, this doesn't matter. For a team that has hard-coded the model into the CI pipeline and tuned the prompt words for weeks, they have to retest every time they change a model. This cost will not appear in the $6.86 per task.

Another is the quality of programming itself. In a single test of the code review tool Bito, Neon scored 79.8%, 0.9 points lower than its own Carbon. Bloomberg reported at the time of Argon’s launch that some employees within Google doubted its actual programming capabilities, and that front-end design was particularly suffering, something Google publicly denied. Neon is based on this foundation to do things quickly and cheaply. Whether the front-end has been improved or not, there is no independent test to answer the question.

The shortcomings of Opus are that it’s expensive, and it’s also counter-intuitive.

The price point has been calculated above.

Counter-intuitively, the Opus 5.5 was trumped on programming benchmarks by its cheaper Sonnet 5.5: 70.6% versus 66.4% by official figures, even though the Sonnet costs half as much. We have also seen before when comparing Sonnet and Opus that the code written by Opus has the lowest bug rate, and it is the best at catching bugs in other people's code.

It's more like a reviewer than a productive writer.

This also explains Google’s table: Neon lost to Claude in the two programming tests. If they were about reading code and finding problems, they were in line with Opus’ strengths. Google didn’t specify in the table what those two tests were, so I can’t fill it in for them.

How will I divide my life?

Leave it to Neon to generate a large amount of code that can be modified and tested. Saving $5.75 per task, running two hundred tasks a day is more than a thousand dollars, enough to run several rounds of tests to filter out what it has compiled.

Leave code review before merging and tasks that require continuous operation in the terminal for a long time to Opus, and turn on caching. Batch is used for nighttime batch processing, and the unit price of Opus is lower than that of Neon.

If you already have Argon or Carbon running on your hand, don’t rush to switch to Neon yet. Once the promotional price expires on November 30, the relative prices of Gemini's own models will be shuffled again, and it won't be too late to wait until the bills come out in the first week of December.

Related comparisons: "Gemini 4 Neon vs. Gemini 4 Carbon: Promotional price expires on November 30, today's price comparison conclusion will be reversed next month", "Claude Sonnet 5.5 and Opus 5.5: Opus ranks last in bug rate when writing code, and Opus wins in catching bugs", "Gemini 4 Carbon, Claude Opus 5.5, GPT-6 Astra: One point ahead Carbon, most people can only look at the list this month》