跪拜 Guibai
← Back to the summary

Gemini 4 Neon Flips the Coding Agent Benchmarks: Faster Than Carbon, Cheaper Than Opus 5.5

Title: Wow, Google revolted, Opus 5.5 got owned again Author: cxuanAI Tags: Backend, Programmer, Artificial Intelligence

Oh, yesterday GPT-6.2 Astra just launched and hit a snag, but after testing it for a day, I found GPT-6.2 Sol works really well. I suggest you guys give it a try.

No time to mourn GPT-6.2 Astra; next up is...

Just now, Google DeepMind officially released Gemini 4 Neon.

Counting it up, this is already the 4th Gemini 4 model Google has released in just 7 weeks - from Argon, Carbon, Hydrogen, to today's Neon. This release pace is faster than my ability to learn this English word and memorize it.

Is Google really going to grind through the entire periodic table?? Google's Gemini 4 Neon official release page

Image source: Google Official Blog, Introducing Gemini 4 Neon.

And it's already available in my subscription plan. My plan is the 5U three-month AI Pro, and it's pretty sweet. image.png

Google's positioning for Neon is very clear: it sits exactly between Gemini 4 Carbon (heaviest, longest-range deep tasks) and Gemini 4 Hydrogen (fastest, high-frequency daily execution), with reasoning capability close to or even surpassing Carbon, paired with much higher speed and lower cost, targeting high-frequency complex programming, knowledge work, and scientific research scenarios.

In other words:

Carbon = Fable Neon = Opus Hydrogen = Sonnet

Google has now turned the tables. It's no longer the days when the Gemini 3.6 / 7 / 8 Flash trio dominated. The only thing still lagging is that Antigravity agent.

Benchmarks directly face-slap Opus 5.5

This time Google went all out, directly comparing Claude Opus 5.5 and yesterday's just-released GPT-6.2 Sol.

Gemini 4 Neon vs Claude Opus 5.5 and GPT-6.2 Sol official benchmark comparison

Image source: Google DeepMind official evaluation table.

Overall capability, in xhigh (Extra High) thinking mode, Neon scored 61 on the Artificial Analysis Intelligence Index, ranking 3rd overall, behind only the larger and slower GPT-6.2 Astra max and Claude Fable 5.2 max, directly surpassing Claude Opus 5.5 and GPT-6.2 Sol, and even scoring 1 point higher than its own heavier Gemini 4 Carbon. It got 70.9% on the Vals Index v2.2, and topped the charts at #1 in both Legal Research Bench and Tax Agent Bench.

Coding Agent is the real headline.

On the third-party independent leaderboard AA, Antigravity CLI running Gemini 4 Neon (xhigh) scored 59, directly taking the #1 spot globally, beating Claude Code running Fable 5.2 and Opus 5.5, and also leaving behind GPT-6.2 Astra running in Codex.

Artificial Analysis Coding Agent Index leaderboard

Image source: Artificial Analysis Coding Agent Index.

Specifically on engineering benchmarks: CursorBench 4.1 got 65.4% (Opus 5.5 was 58.1%); on SWE-sweep, in the generic mini-SWE harness it led with 12.8% (Opus 5.5 only 3.6%), and switching to its own Antigravity harness it jumped over threefold to 39.0% (Opus 5.5 was 10.1%, GPT-6.2 Sol only 8.3%).

However, objectively speaking, on Terminal-Bench 4.1 and FrontierCode 2.0, Opus 5.5 still edges out slightly.

Long context and agent workflow gap is even more absurd.

On GraphWalks with 256K to 1M tokens, Neon shot up to 95.2%, a full 28.4 percentage points higher than Opus 5.5, and a huge jump from last month's Argon at 84.2%. Zapier's AutomationBench and the simulated business long-horizon test Vending-Bench 2 both took #1; the Arena's newly launched Agent Arena Index (covering 4,200+ real agent sessions) debuted globally at #2 (1336 points), with Steerability (mid-task obedience and correction ability) also ranking #2.

And officially, Neon recorded the lowest hallucination rate across the entire family in Antigravity internal tests, and achieved a 0% attack success rate on the Gray Swan IPI indirect prompt injection test with real-time protection enabled.

In terms of speed and per-task cost:

According to AA's measurements, Neon's output speed is about 110 tokens/s, nearly twice that of Carbon, and the average per-task time is about 40% less than Carbon.

Running the same Coding Agent benchmark suite, Antigravity + Neon (xhigh) averaged only $6.86 per task, while Claude Opus 5.5 (max) cost $12.61 and Claude Fable 5.2 (xhigh) was as high as $5.47, saving nearly half directly. Even looking at the default High setting, Neon's per-task cost is 32% cheaper than Carbon.

Artificial Analysis per-task cost vs intelligence index Pareto curve

Here's another very intuitive chart. On the AA-Briefcase presentation quality leaderboard, Neon shot to 1771 Elo, ranking #1 globally. Artificial Analysis published a side-by-side comparison of a slide generated by Neon and Opus 5.5 on the same business analysis prompt:

AA-Briefcase presentation test comparison: Gemini 4 Neon ($4.03) vs Claude Opus 5.5 ($30.06)

Image source: Artificial Analysis, Gemini 4 Neon measurement report.

The top slide above cost Neon $4.03; the bottom slide cost Opus 5.5 $30.06.

Orchestrator (Neon): let Neon think, Hydrogen work

Also launching alongside Neon this time is the Orchestrator (Neon) I mentioned in the earlier screenshot.

Many people would initially think this is just a front-end router, but it's not. According to the Antigravity official documentation, it's a relay between two models within the same task round:

The official demo videos and screenshots below show more intuitively what Neon can do with Antigravity's /swarm.

Demo 1: Use /swarm to hand-build a 3D neon golf game, AI plays as Player 2 itself

The first is Google Antigravity's flagship case: NEONPUTT '84.

Gemini 4 Neon spun up an 8-agent /swarm in Antigravity to build in parallel - from project scaffolding, a 240Hz fixed timestep physics engine, Shaders, 8 cyber-neon fairways, to UI and acceptance testing. The Swarm Graph decomposed tasks and agents @mentioned each other to collaborate.

NeonPutt '84 Demo: 8 agents collaborating in real-time through Swarm Graph to decompose tasks

After all 8 fairways were done, the agents launched localhost:5173 in the built-in browser and self-checked each level's rendering, physics collisions, and interaction feel.

The most absurd part is the second half: the user simply said "I'll play Player 1, you only play Player 2, watch me take this shot." Then Antigravity actually took over the built-in browser, looked at the fairway layout, aimed the club, and played a two-player mini-golf round with a real human.

NeonPutt '84 Demo: Antigravity takes over the built-in browser as Player 2 to play against a human

Full official demo video:

![[gemini4-neon-neonputt-demo.mp4]]

From writing the game, self-testing and acceptance, to playing against you as a second player, the whole loop happens in the same window.

Demo 2: Blender MCP recreates the "Periodic Table" physics lab

The second one continues the Blender MCP route that impressed with Argon last time, but this time Google plays on its own naming joke.

Gemini 4 Neon used /swarm in Antigravity to drive Blender MCP, building a photorealistic 3D physical chemistry lab bench, bringing all the Gemini 4 family element names to life: on the left are Carbon and Hydrogen represented by Bunsen burner flames, and on the right are Argon and Neon represented by high-pressure gas discharge tubes.

Gemini Elements 3D Demo: Blender MCP rendered Bunsen burner flames and high-pressure gas discharge tubes

If you weren't told this was made by the model through repeated Blender MCP rendering and checks, you'd probably mistake it for a real recorded experiment video.

Demo 3: Zero preset assets, hand-coded first-person WebGL2 ski simulator on Mont Blanc

The engineering weight of this demo will make anyone who knows graphics and frontend get goosebumps.

Gemini 4 Neon directly used real elevation data of Mont Blanc and Chamonix, and from scratch in Antigravity wrote a first-person WebGL2 freestyle ski simulator.

Chamonix Ski Simulator Demo: WebGL2 first-person skiing footage based on real Mont Blanc terrain

Full video below.

Demo 4: Digging through historical archives, reviving Turing's 1936 study 1:1

The fourth is a historical scene restoration, and the quality immediately feels different.

Gemini 4 Neon, again through /swarm, searched historical archive photos and turned Alan Turing's 1936 study at King's College, Cambridge, into an interactive 3D space.

Turing Machine 1936 Demo: King's College Room X.8 study with a working universal Turing machine on the desk

From the King's College chapel outside, to the fireplace, bookshelves, green banker's lamp, vintage typewriter, and blackboard covered in state transition tables, even the universal Turing machine on the desk is actually runnable inside.

When Argon turned medical materials into a 3D anatomy tool last time, I thought turning dry historical materials into walk-in, operable interactive spaces was especially interesting. Now they can even restore Turing's 1936 study directly.

Demo 5: Using pure Google Sheets formulas to run computer vision and 3D relighting

The finale, this one is absolutely insane.

Neon's Agent Swarm stuffed an image into a Google Sheets with 100,000 cells: no embedded image used, not a line of Apps Script written, purely with over 500,000 live spreadsheet formulas, it ran edge detection, 3D depth map mapping, and supported 6 interactive scene relighting modes right inside the spreadsheet.

Sheets CV Lab 100K Demo: Running edge detection and 3D relighting with 500,000 Google Sheets live formulas

This creativity and engineering execution is genuinely something.

Pricing

Finally, the API pricing everyone cares about. Gemini 4 Neon vs Claude Opus 5.5, Gemini 4 Carbon, GPT-6.2 Sol, GPT-6.2 Astra standard API price comparison

Neon's official standard price is set at $3 / million input tokens, $15 / million output tokens, with cache hit input at 0.5折 (i.e., 0.5折 means 5% discount, so $0.15 / million tokens), and a flat $0.15 / million tokens for cache-hit input. Also, 1M context is a flat price, unlike the GPT-6.2 series which adds surcharges beyond 272K.

A couple of days ago I wrote that with A (Anthropic) mass-banning accounts and GPT models' reputation falling apart, Google might be the answer at this stage.

Right after I said that, Google served up Gemini 4 Neon: double speed, another quarter cut in price, directly included in the AI Pro plan, and killed Opus 5.5.

In my situation where my Claude Max plan hit KYC verification and got locked, Apple refused a refund, A couldn't find a single human to talk to let alone refund me, so I had to file complaints with the California government and the BBB.

For the coming period, my main model is likely going to officially switch to Neon; GPT-6.2 Sol can be my assistant.