Article title: Just now, Gemini 4 Neon surprise-launched, killing Opus 5.5 on cost-effectiveness overnight
Body:

Just now, Gemini 4 Neon surprise-launched, killing Opus 5.5 on cost-effectiveness overnight
2026-11-19 07:56
You have to admire Gemini's courage in picking a date to release news. Last night, "GTA 6" had just unlocked and gone live, wiping the global internet trending searches clean, yet Brother Pichai today still braved the sky full of gaming posts and forcibly pulled another new model out of the periodic table.

Just now, only three weeks after the release of the previous flagship model Gemini 4 Carbon, Google turned around and served up a brand-new main flagship large model, Gemini 4 Neon—that is, element No. 10 in the periodic table, the noble gas "neon."

What most reassured developers was that Google did not pull the queue-and-apply futures gimmick again this time. On launch day, it went fully live in stock simultaneously on Google AI Studio, the Gemini App, the enterprise API, and its own coding environment Google Antigravity.
If Carbon from three weeks ago was the flagship model meant to take on ultra-long-horizon complex tasks, then Neon, which takes the stage today, is the volume-driving cost-effectiveness mainstay Google has truly sent out to fight at close quarters with Claude Opus 5.5 and GPT-6.2 Sol.
It not only scored a high 61 points on the Artificial Analysis composite intelligence leaderboard, but also pushed pricing directly down to $3 per million input tokens (cached input as low as $0.15), $15 for output, while standardly offering a 1 million context window.

Even Rockstar knows to deliver on time, so Google naturally felt embarrassed to keep selling futures.
The Gemini 4 family seating chart: why Neon gets the C position
Before discussing model capabilities, one cannot help but first complain about Google's now-dizzying chemical nomenclature. In the past, when naming large models, everyone at most permuted and combined Pro, Max, Ultra, and Flash. Although clichéd, users could tell at a glance who was stronger and who was weaker.
ChatGPT is the astronomy department, Claude is the literature/poetry department, and by the Gemini 4 generation, Google simply plunged headlong into the periodic table, successively pulling out No. 1 "Hydrogen," No. 18 "Argon," No. 6 "Carbon," and today's No. 10 "Neon."
It really is a complete comprehensive exam paper spanning math, physics, chemistry, and the liberal arts.
Even a researcher from Google DeepMind itself, 姚顺宇, joked on X today, saying that using chemical elements as codenames is simply a genius idea. Every time a codename leaks, watching netizens be completely baffled and unable to figure out which tier the model is in or who it should be benchmarked against is just too interesting.

To prevent everyone from being thoroughly confused, Google finally clarified the Gemini 4 family seating chart in its official chart this time.
From the officially announced family positioning table, the main lineup is split into clear left, center, and right tiers.
On the far right, Hydrogen focuses on everyday high-frequency programming and lightweight Agent workflows, with an AA intelligence index of 54 points and only $1.5 and $7.5 per million input and output tokens, open to everyone.
On the far left, Carbon specializes in the hardest ultra-long-horizon large projects and deep research, with an AA intelligence index of 60 points. Standard pricing is $4 per million input and $20 output (promotional price of $2 and $10 before November 30), mainly aimed at high-end Ultra subscribers.

Neon, released today, lands right in the golden sweet spot in the exact middle, focusing on software engineering, knowledge work, and scientific reasoning, and users on both Pro and Ultra tiers can use it directly.
Interestingly, because Neon offers one more extreme thinking level called "xhigh" than Carbon, its AA intelligence index actually rose to 61 points, overturning its own big brother Carbon by a 1-point margin, yet its everyday pricing of $3 input and $15 output is a quarter cheaper than Carbon's standard price.
Perhaps fearing that the wealthy Ultra subscribers paying high prices would feel unbalanced, Google today also casually stuffed a big Easter egg to eligible Ultra users, officially opening the new Project Genie experience and the world model Genie 4.

Users do not need to configure any top-tier graphics card; they can directly summon the Project Genie agent in the browser to generate and walk into interactive 3D virtual worlds, and can even directly remix other people's worlds in the community gallery.
If you haven't had time to play "GTA 6" yet, you might as well try hand-building a cyber playground in Genie 4.
Facing off against Claude and GPT: Neon challenges its two great rivals
Since it dares to debut in the family's C position, Neon's report card this time is naturally aimed at its peers' vital points.
In the official all-dimension benchmark chart, Google directly pulled Neon, rival Claude Opus 5.5, and GPT-6.2 Sol onto the same table for close-up comparison. Overall, the three have wins and losses against each other, but in its strong suits Neon bites extremely hard.
On traditional command line and frontier code sub-items (Terminal-bench 4.1 and FrontierCode 2.0), Opus 5.5 still maintains a slight lead; in dense chart reading (Chartography), GPT-6.2 Sol wins by a small margin.

But once it switches to composite intelligence (AA Index 61 points vs 59 points vs 57 points), Agent programming (CursorBench 65.4%, and under its own Antigravity framework SWE-sweep even soared directly to 39.0%), mathematical and scientific reasoning, and cybersecurity vulnerability discovery, Neon's advantage is already quite obvious.
The most exaggerated is the GraphWalks test, which examines 256K to 1M ultra-long context retrieval. Neon directly ran out a score of 95.2%, leaving Opus 5.5's 66.8% and GPT-6.2 Sol's 80.1% a full body length behind.

To prove that this model can really do heavy engineering work, the four demo demos Google officially ran with Antigravity are also quite wild in imagination.
For example, it built a 3D chemistry lab with a Bunsen burner and a high-voltage discharge tube in Blender to play on its own model's name.

Another example is a 1:1 restoration, based on archival photos, of Turing's 3D study at King's College, Cambridge in 1936. It could even write 10,200 lines of custom Shader code from scratch to develop a WebGL2 Mont Blanc skiing game using 120Hz physics simulation, complete with dynamic snow tracks, avalanches, and spatial audio.

And the most interesting is surely Google having a group of Neon agents turn a static photo into a Google Sheet with 100,000 cells.
Without inserting a single image or writing a single line of script code, the AI managed, purely relying on more than 500,000 real-time spreadsheet function formulas, to run edge detection, 3D depth mapping, and six interactive lighting modes inside the spreadsheet—comparable to using an Excel sheet as a 3D graphics card.

Neon charges into the coding battlefield, and the most lethal thing is the price
Because this time it is available in stock, quite a few netizens have already played with it in advance or officially.
AI blogger Chubby soaked in Antigravity with Neon for a day and generated 17 interactive demos covering shooter games and 3D worlds in one go.

Values written like {{NAME_ab12cd34}} are placeholders for real values. Use them exactly as written, in replies and in tool calls alike: they are filled in with the real values before anything runs. Don't ask to verify them, and don't replace them with examples.
Just Now, Gemini 4 Neon Launches in a Surprise Raid, Slaying Opus 5.5 on Value Overnight
While testing a destructible voxel arena called “Terra Breaker,” he originally only asked for destruction and reset functions, but Neon proactively filled in three types of buildings and three destruction tools for him: a cannon blast could send a large swath of blocks flying, a chisel strike only knocked off a few fragments, and it ran steadily at a full 60 frames in the browser.

And in another game about flying a jet through sunset canyon obstacles, Neon even, without anyone asking, casually wrote an extra autopilot demo mode for the plane.
Not only does it not cut corners lazily, it even proactively gives itself extra scenes.
This also corroborates a detail mentioned in Google’s safety report: Neon’s cheating rate in Terminal-Bench 4.0 was pinned firmly at 0.0%, and in internal coding anti-laziness tests, its rate of taking shortcuts also dropped to a historic low of 2.45%.

As for what everyone cares about most—who is actually stronger in real tests against its nemesis Claude Opus 5.5—the blogger JAZII set up a very intuitive head-to-head on the same task.
With both sides running at the highest reasoning setting, Neon took 20 minutes to finish the task, with a bill of $3.5; Opus 5.5 took 18 minutes, with a bill as high as $7.2.

The final feel was clear: Opus 5.5 remains the slightly superior benchmark at the absolute upper limit, but Neon delivered the most refined output in the entire Gemini 4 family to date, and cut the cost of a single task by more than half outright.

As AI tech company Glean put it, Neon can deliver results of equivalent quality at 3 times the speed and 60% of the cost of Carbon, with overall per-task cost 50% to 80% lower than top competitors on the market.
What’s more interesting is the timing of Google’s move.
Right now, as Anthropic is busily preparing for the key window of its IPO, OpenAI and Google have, over the past two days, been taking turns with remarkable tacit understanding—each trying to outdo the other in dumping low-price mainstay models into programmers’ and enterprise workflow heartlands, and it’s hard to say there’s no intention of casually needling their old rival and sticking in a knife.
Once the bell of the capital market rings, it will probably get even livelier.
But for those of us wage earners who grind away at code and spreadsheets every day, gods fighting is always a good thing: as long as the big companies are willing to roll out top-tier off-the-shelf models at half price, everyone will happily use Neon to save on compute costs while getting off work early to go home and keep grinding GTA 6.
And at Google’s pace of turning out a model every three weeks, there aren’t many memorable chemical elements left for them.