Google Releases Gemini 4 Neon, Highlighting Coding and Cost-Effectiveness
Google has released Gemini 4 Neon for software engineering, knowledge work, and scientific reasoning, making it available the same day through Google AI Studio, the Gemini App, enterprise APIs, and Google Antigravity. The model offers a 1-million-token context window and is priced at $3 per million input tokens and $15 per million output tokens. It scored 61 on Artificial Analysis’s overall intelligence index. The article states that it achieved 95.2% on GraphWalks and summarizes comparisons with Claude Opus 5.5 and GPT-6.2 Sol in coding, reasoning, and cost.
Why read this
AI score 88
The material summarizes Gemini 4 Neon’s pricing, context window, and multiple comparative benchmarks, providing a basis for evaluating the trade-offs between cost and capabilities in programming workflows.
Topics
Body · AI translation
You have to admire Gemini’s courage in choosing the day to make news. Last night, GTA 6 had barely unlocked and launched before it completely dominated trending searches across the global internet, yet today Sundar Pichai still pushed through the deluge of gaming posts and somehow pulled another new model out of the periodic table.
Just moments ago, only three weeks after the previous flagship model, Gemini 4 Carbon, was released, Google unveiled its brand-new mainstay large model, Gemini 4 Neon—named after neon, the rare gas with atomic number 10 on the periodic table.
What will come as the biggest relief to developers is that Google has dispensed with the old futures-style game of making people apply and wait in line. On launch day, the model became fully and immediately available through Google AI Studio, the Gemini App, enterprise APIs, and Google’s own coding environment, Google Antigravity.
First, here are the key takeaways.
If Carbon, released three weeks ago, is the heavyweight model designed to tackle complex, long-horizon tasks, then Neon is the high-volume, cost-effective workhorse Google is truly fielding for a head-to-head fight with Claude Opus 5.5 and GPT-6.2 Sol.
Not only did it score an impressive 61 on the Artificial Analysis Intelligence Index, but its pricing has also been pushed down to $3 per million input tokens—with cached input costing as little as $0.15—and $15 per million output tokens, while a 1-million-token context window comes standard.
Even Rockstar managed to deliver on time, so Google could hardly keep announcing products that were not yet available.
Ranking the Gemini 4 family: Why does Neon deserve center stage?
Before discussing the model’s capabilities, it is impossible not to complain about Google’s increasingly bewildering chemical naming scheme. Large models used to be named with some combination of Pro, Max, Ultra, and Flash. It may have been clichéd, but users could tell at a glance which was stronger and which was weaker.
Now ChatGPT has gone with astronomy, Claude with literature and poetry, while in the Gemini 4 era Google has plunged straight into the periodic table, successively unveiling No. 1 “Hydrogen,” No. 18 “Argon,” No. 6 “Carbon,” and today’s No. 10 “Neon.”
At this point, they have practically assembled a comprehensive exam covering math, physics, chemistry, and the humanities.
Even Google DeepMind researcher Shunyu Yao joked on X today that using chemical elements as codenames was a stroke of genius. Every time a codename leaks, it is enormously entertaining to watch people online scratch their heads, completely unable to figure out what tier the model belongs to or which rival it is supposed to compete with.
To keep everyone from becoming completely confused, Google has finally clarified the hierarchy of the Gemini 4 family in an official chart this time.
According to the officially published family positioning table, the core lineup is divided into three clearly defined tiers: left, center, and right.
On the far right, Hydrogen focuses on frequent everyday coding and lightweight Agent workflows. It scores 54 on the AA Intelligence Index, costs just $1.50 per million input tokens and $7.50 per million output tokens, and is available to everyone.
On the far left, Carbon is built to chew through the toughest long-horizon projects and deep research. It scores 60 on the AA Intelligence Index and has standard pricing of $4 per million input tokens and $20 per million output tokens, with promotional prices of $2 and $10 through November 30. It is aimed primarily at advanced Ultra subscribers.
Neon, released today, sits squarely in the golden sweet spot in the middle, focusing on software engineering, knowledge work, and scientific reasoning. Both Pro and Ultra users can access it directly.
Interestingly, because Neon offers an additional maximum-reasoning setting called “xhigh” that Carbon lacks, it actually reached 61 on the AA Intelligence Index, edging out its older sibling Carbon by one point. Yet its regular pricing of $3 per million input tokens and $15 per million output tokens is still 25% lower than Carbon’s standard rates.
Perhaps concerned that affluent Ultra subscribers paying premium prices might feel shortchanged, Google also slipped eligible Ultra users a major bonus today, officially opening access to the new Project Genie experience and the Genie 4 world model.
Users do not need a top-of-the-line graphics card. They can summon the Project Genie agent directly in a browser to generate and enter an interactive 3D virtual world, and can even remix other people’s worlds from the community gallery.
If you have not had a chance to play GTA 6 yet, you might as well try building a cyber amusement park by hand in Genie 4.
Taking on Claude and GPT head-on: Neon challenges its two archrivals
Since Neon is bold enough to debut in the center position of the family lineup, the results it has published are naturally aimed straight at its competitors’ weak spots.
In the full benchmark chart officially released, Google directly put Neon at the same table as its archrivals Claude Opus 5.5 and GPT-6.2 Sol for a head-to-head comparison. Overall, each of the three scored its share of wins, but Neon was exceptionally competitive in its areas of strength.
On the traditional command-line and frontier coding subtests—Terminal-bench 4.1 and FrontierCode 2.0—Opus 5.5 still maintained a slight lead, while GPT-6.2 Sol narrowly prevailed in dense chart reading (Chartography).
But when it came to general intelligence (AA Index scores of 61 vs. 59 vs. 57), agentic programming (65.4% on CursorBench, with SWE-sweep soaring to 39.0% under Google’s own Antigravity framework), mathematical and scientific reasoning, and cybersecurity vulnerability discovery, Neon’s advantage was already quite apparent.
Most strikingly, on the GraphWalks test, which evaluates retrieval across ultra-long contexts ranging from 256K to 1M, Neon scored 95.2%, pulling well ahead of Opus 5.5’s 66.8% and GPT-6.2 Sol’s 80.1%.
To prove that the model can genuinely handle heavy engineering workloads, Google also produced four remarkably imaginative demos using Antigravity.
In one, it built a 3D chemistry laboratory in Blender, complete with a Bunsen burner and a high-voltage discharge tube, as a play on the model’s name.
In another, it recreated Turing’s 1936 study at King’s College, Cambridge, in 3D and at 1:1 scale based on archival photographs. It could even write 10,200 lines of custom Shader code from scratch to develop a WebGL2 Mont Blanc skiing game with 120Hz physics simulation, including dynamic ski tracks, avalanches, and spatial audio.
Most interesting of all, however, Google had a group of Neon agents turn a static photograph into a Google Sheet containing 100,000 cells.
Without inserting a single image or writing a line of script code, the AI relied entirely on more than 500,000 live spreadsheet formulas to implement edge detection, 3D depth mapping, and six interactive lighting modes inside the spreadsheet—essentially using Excel as a 3D graphics card.
Neon enters the coding battlefield, with price as its deadliest weapon
Because the model was available immediately this time, many users had already tried it through early access or upon its official release.
AI blogger Chubby spent a day using Neon in Antigravity and generated 17 interactive demos in one go, spanning shooting games and 3D worlds.
While testing a destructible voxel arena called Terra Breaker, he had initially asked only for destruction and reset features. Neon instead proactively added three types of structures and three destruction tools. Firing the cannon could blast away a large swath of blocks, while striking with the chisel removed only a few fragments, and the game held a steady 60 frames per second in the browser.
In another game about flying a jet through obstacles in a canyon at sunset, Neon even took it upon itself to add an autopilot demo mode, despite no one asking for one.
Not only did it avoid cutting corners, it even proactively gave itself extra work.
This also corroborates a detail mentioned in Google’s safety report: Neon’s cheating rate on Terminal-Bench 4.0 was held firmly at 0.0%, while its rate of taking shortcuts in an internal anti-laziness coding test fell to an all-time low of 2.45%.
As for the question everyone cares about most—whether it is actually stronger than its archrival Claude Opus 5.5—blogger JAZII conducted a highly intuitive head-to-head test using the same task.
With both models set to their highest reasoning level, Neon completed the task in 20 minutes and cost $3.50, while Opus 5.5 took 18 minutes and ran up a bill of $7.20.
The final impression was clear: Opus 5.5 remained the benchmark with a slight edge at the absolute performance ceiling, but Neon delivered the most refined output yet from the entire Gemini 4 family while cutting the cost per task by more than half.
As AI technology company Glean put it, Neon can deliver results of equivalent quality at three times Carbon’s speed and 60% of its cost, with total per-task expenses 50% to 80% lower than those of the leading competing models on the market.
Even more interesting was the timing of Google’s move.
Anthropic is currently in the crucial window of intensive preparations for an IPO, yet OpenAI and Google have shown remarkable coordination over the past two days, taking turns launching aggressively priced flagship models deep into the territory of programmers and enterprise workflows. It is hard to believe they were not also looking to put pressure on their longtime rival and land another blow while they were at it.
Once the opening bell rings in the capital markets, things will probably become even livelier.
Still, for those of us who spend every day wrestling with code and spreadsheets, clashes between the giants are ultimately a good thing. As long as the major companies are willing to push their top, readily available models onto the market at half price, everyone will happily use Neon to save on compute costs, finish work earlier, and go home to continue grinding GTA 6.
At the rate Google is releasing models, it is quickly running out of memorable chemical elements to name them after.
This article originally appeared on the WeChat public account “APPSO” and was written by APPSO, Discovering Tomorrow’s Products. It is republished by 36Kr with authorization.
AI Trends