Astra 6.2 tested15 minOpenAI's new GPT-6.2 Astra and Sol launched on November 17 with a 20% price cut and a wave of harsh early reviews. We walked through a full hands-on gauntlet, from a two-piano duel and an Unreal Engine sword fight to a hidden frog and six brain scans, to see where the new model actually lands.Read now →

Gemini 4 Neon Played a Live Piano Duet and Rebuilt a Villa in 21 Minutes. Then It Couldn't Find a Frog
Google's 'middle' model, Gemini 4 Neon, went through ten hard hands-on tests: a two-piano duet, video effects reverse engineering, an animated biology lesson, a full song, an Unreal Engine sword fight, an Airbnb rebuilt in Blender, brain scans and more. It was faster than its rivals almost everywhere. It was not better everywhere.
Benchmarks tell you how a model does on a test someone designed. They don't tell you what happens when you hand it a messy, real project and walk away.
So when Google released Gemini 4 Neon on November 18, 2026, and early leaderboards put it ahead of models that cost far more, the obvious next question was: what does it actually do with hard work?
Neon has now been through a full gauntlet of ten demanding hands-on tasks, the same kind of run GPT-6.2 Astra faced a day earlier. Two pianos at once. A video's special effects, reverse engineered. A song with AI vocals. A 3D sword fight built from nothing. A real vacation rental rebuilt in Blender. Six brain scans.
The short version: Neon was fast everywhere, brilliant in a few places, and embarrassing in at least one. Here's every test, what happened, and what it tells you about whether to use it.
If you want the background on what Neon is and why its pricing matters, start with our explainer on how Neon beat Carbon. This piece is about what it does when you actually put it to work.
The setup
A quick orientation before the tests:
- Neon sits in the middle of Google's new Gemini 4 family, between the flagship Carbon and the lighter Hydrogen. It's not supposed to be Google's smartest model.
- It has a 1 million token context window and offers thinking levels up to "xhigh," according to Google's model card.
- In the API it costs $3 per million input tokens and $15 per million output tokens.
- In the consumer app, Google says Neon is available to Pro and Ultra subscribers, while Carbon needs the top Ultra plan.
For the tough multi-file, multi-step tasks, the tests ran inside Antigravity, Google's own agent harness, using Neon at its highest "Extra" setting with a permissions mode that let it work without stopping to ask. Simpler questions went through the regular Gemini chat.
One more thing to watch throughout: time. Neon's speed turned out to be its most consistent advantage.
Test 1: a live piano duet, played on two browsers at once
Top models can already compose a short piano piece and play it live on a web keyboard. This test doubled the difficulty.
The prompt asked Neon to:
- open two separate browsers, each with a different online piano,
- compose a fast, roughly one-minute duet in the style of Chopin for two players,
- and perform it live, one player per piano, with both parts in sync.
Neon thought for about 10 minutes, mostly testing keys and figuring out how each piano's interface worked. Then it played.
Verdict: a clear win. The duet came off in sync, and the result was judged clearly better than GPT-6.2 Astra's attempt at the same prompt, done in a fraction of the time. Coordinating two different interfaces at once is a genuine computer-use challenge, and this is where Neon looked most impressive.
Test 2: reverse engineering a video's effects
Next, Neon got a short video packed with editing tricks:
- a very fast zoom in and out,
- a random camera-shake effect,
- a Polaroid-photo style transition,
- a chromatic aberration "glitch,"
- and a cross-zoom into the final frame.
The task: figure out each effect from the video alone, then produce a new video that alternates between two different images using the same transitions.
It worked for 17 minutes. The result reproduced the transitions and effects almost perfectly. The only visible miss was the last one, where Neon used something closer to a radial zoom blur instead of the original cross-zoom.
Verdict: very strong. If you edit video and want to copy a look you've seen, this is a practical use case today.
Test 3: an animated biology lesson, from scratch
The prompt asked for an immersive explainer that flies through an animal cell, introducing each organelle, with no external animation tools or video generators, and narration from Google's own text-to-speech.
The first pass took 24 minutes, noticeably quicker than Opus or GPT-6.2 on similar work. It looked basic, so a follow-up asked for much more detail and accuracy.
The final video walked through the plasma membrane, cytoplasm and cytoskeleton, the nucleus and nucleolus, rough and smooth endoplasmic reticulum, ribosomes, the Golgi apparatus, mitochondria, lysosomes, peroxisomes and the centrosome. The narration was solid.
The visuals were not. Parts still looked like simple shapes, and some shapes overlapped awkwardly at the edges.
Verdict: a tie with GPT-6.2 Astra. Good script, average animation.
Test 4: write, sing and produce a full song
This one stacked several hard jobs:
- compose an upbeat future-bass track, about 3 minutes long,
- write English lyrics about "accelerating into the singularity" for a female voice,
- generate the vocals in ACE Studio, an AI singing platform,
- build the instrumental in a real digital audio workstation, using any plugins or downloaded samples,
- add effects and "ear candy," then mix and master it.
After 15 minutes it had a decent draft. A follow-up asked it to make the vocals sit better with the instrumental and push the chorus louder.
The finished project was surprisingly thorough: a lead vocal, harmony vocals, and an instrumental with multiple synths, an electric piano, sparkles, leads, pads, percussion, effects, chords and even vocal chops. The singing came out a bit too high-pitched.
Verdict: good, and better than GPT-6.2 Astra's attempt at the same task.
Test 5: a 3D sword fight in Unreal Engine (built from nothing)
This was designed to be brutal. No existing character models allowed. Neon had to:
- create a realistic heroine from reference images, including hair and clothing movement,
- create a ninja antagonist,
- rig and animate both so they could run, dodge, jump and slash (with permission to pull animations from Mixamo, a free animation library),
- build an ancient Chinese palace interior,
- do the modeling in Blender and render the game in Unreal Engine,
- and make it look like a big-budget game.
It also ran a built-in quality loop: a separate agent took screenshots, scored them, and sent the work back if the score was under 8.5 out of 10, for up to four rounds.
Neon worked for about 40 minutes. After four rounds, its own critic gave it 4.5 out of 10. Two more rounds of "make the characters more faithful" improved things only slightly. The game then threw an error, which Neon fixed when the error was pasted back in.
The playable result: a decent palace, but characters with broken joints, a heroine not holding her sword properly, and rigging and animation that were "all messed up."
Verdict: a fail, and GPT-6.2 Astra did better here. Creating organic 3D characters from scratch remains one of the hardest things you can ask any AI to do.
Test 6: rebuild a real vacation rental in Blender
Here's where Neon's speed really showed.
It got photos from a real rental listing: a living room with sofas and a dining table facing a pool, a bedroom facing a circular pool, a single photo of the bathroom with no clue where it sits, exterior shots, and a pool with an unusual slanted wall. The task: reconstruct the whole building and interior in Blender, faithfully, and program a camera to fly through it like a virtual tour.
Neon finished in about 21 minutes. Other frontier models, including GPT and Opus, took over an hour on the same prompt.
One round of feedback asked for more realism, especially the sofa and bed pillows. The final version:
- got the spatial layout right, including the exterior,
- produced convincing sofas and pillows,
- nailed the dining table, including its items and textures,
- and rebuilt the bedroom, down to the ceiling lights.
Verdict: excellent, especially for the price tier. For architecture visualization and 3D mockups, this is Neon's best showing.
Test 7: find the hidden frog
A classic visual test. An image contains a frog camouflaged somewhere in a busy scene. Neon wasn't told there was a frog, only asked: is there any animal in this image? If so, identify and circle it.
It didn't find it. It said there was no animal.
One frustration: the chat interface didn't show its reasoning, so there was no way to see how it searched.
There's a silver lining, though, and it leads to one of Neon's real strengths. More on that below.
Test 8: six brain scans
Six MRI images, each containing a very early-stage brain tumor, some hard even for human experts to spot. Neon had to identify the tumor type in each, if any.
Its answers:
- Top left: suprasellar meningioma. Correct.
- Top middle: pineal region tumor. Wrong.
- Top right: multiple meningiomas. Partially right (half credit).
- Bottom left: no tumor. Wrong.
- Bottom middle and bottom right: craniopharyngioma. Both wrong.
Score: 1.5 out of 6. That sounds bad, but context matters. GPT-6.2 Astra scored zero on the same set. Claude Fable 5.2, considered the strongest model on this test, managed two.
Verdict: relatively good, absolutely not a diagnostic tool. Nobody should use any chatbot to read their own scans.
Test 9: deep research on leukemia
The prompt asked for the molecular drivers of a specific type of leukemia, how targeted therapy evolved, and a comparison of resistance mechanisms and survival outcomes from recent long-term studies.
Neon produced a long answer with tables comparing drugs, their safety and efficacy, resistance mechanisms and survival outcomes.
Two problems:
- No citations. For a medical research answer, that makes fact-checking almost impossible.
- Too verbose. Long, dense paragraphs that were hard to read and slow to get to the point.
Verdict: competent but frustrating. All top models can do deep research; this one's style simply wasn't preferred over GPT's.
Test 10: invent new ways to deal with plastic waste
The last test asked for three novel, scalable solutions to plastic waste that would be cost-competitive with virgin plastic and work on mixed, contaminated plastic.
Neon proposed:
- Inline reactive vitrification: skip sorting entirely and feed mixed plastics into an industrial twin-screw reactive extruder with a small dose of catalysts and cheap alkaline scavengers.
- An electrified supercritical-water refinery with in-situ dechlorination.
- A "tandem abiotic-biotic biological funneling cascade," combining chemical breakdown with biological processing.
Each came with an explanation, why it was novel, the main challenges, near-term feasibility, a summary table and a flowchart.
Verdict: ambitious, but jargon-heavy. There's no right answer to an invention prompt, and the ideas were at least structured. The writing again leaned on complicated terms where plain language would have helped.
The scorecard
| Test | Result | Versus GPT-6.2 Astra |
|---|---|---|
| Two-piano duet | Clear win | Better |
| Video effects reverse engineering | Near perfect | Not compared |
| Animated cell lesson | Average visuals | Tied |
| Song production | Good, vocals too high | Better |
| Unreal Engine sword fight | Fail (4.5/10) | Worse |
| Blender villa | Excellent, 21 minutes | Much faster |
| Hidden frog | Missed | Not compared |
| Six brain scans | 1.5/6 | Better (Astra 0) |
| Deep research | Verbose, no citations | Worse |
| Plastic invention | Ambitious, jargon-heavy | Not compared |
And the usage: across all of these tests, run on Google's Ultra plan, Neon used only about 11% of the weekly usage limit. That's remarkably efficient for this much heavy agent work.
What the benchmarks add
The hands-on picture lines up with the published numbers in interesting ways.
Long context is a standout. On GraphWalks, which tests reasoning over enormous inputs of 256,000 to 1 million tokens, Google's model card reports Neon at 95.2%, against 66.8% for Claude Opus 5.5 and 80.1% for GPT-6.2 Sol.
It beats its bigger sibling on the main index. On the Artificial Analysis Intelligence Index (v5.0.1), the model card lists Neon at 61, ahead of Opus 5.5 (59) and GPT-6.2 Sol (57). Third-party comparisons place it just behind GPT-6.2 Astra (62) and Claude Fable 5.2.
Independent leaderboards like it. BenchLeader's tracker lists Neon at 87.4% on LiveBench 2.0 and ranks it #1 on LiveBench's research category.
Speed vs. cost is a real trade-off. One detailed analysis of Artificial Analysis data shows Neon at "high" scoring 60 for about $3.74 per task in 14.8 minutes, while "xhigh" adds just one point but raises cost by roughly 72% and runtime to 21.5 minutes. Astra at max scored 62 for about $4.00. So Neon is far cheaper per task than Opus 5.5 or Fable, but not cheaper than Astra at the top setting.
The hallucination story (and the frog)
This is the most important number for everyday users.
Artificial Analysis's Omniscience benchmark measures two things: how often a model gets factual questions right, and how often it makes something up instead of admitting it doesn't know.
Gemini models are notably more willing to say "I don't know." Data reported on November 19 shows:
- Gemini 4 Neon (high): about 59% accurate, 13.9% hallucination rate.
- Claude Opus 5.5 (max): about 66% accurate, 50.3% hallucination rate.
- GPT-6.2 Astra (max): about 69% accurate, 59.1% hallucination rate.
- Claude Fable 5.2 (max): about 79% accurate, 59.3% hallucination rate.
In plain English: the Claude and GPT models know more, but when they don't know, they're far more likely to bluff. Neon knows a bit less and bluffs far less.
That reframes the frog. Neon didn't invent an animal or circle a random leaf. It said it saw nothing. Wrong, but honest. For research, legal or financial work, that trait can matter more than a few points of raw accuracy.
Practical takeaways
- Use Neon for speed-sensitive agent work: 3D mockups, browser automation, video effects, anything where a rival model would grind for an hour.
- Use it for huge documents. The long-context numbers are the best in its class.
- Prefer it when bluffing is costly. Its low hallucination rate is a genuine differentiator.
- Don't use it for organic 3D characters or as your only deep-research tool. Ask for citations explicitly, and ask for concise answers.
- Start at "high," not "xhigh." The extra level adds little score for a lot more time and money.
- Check your plan. Neon is available to Pro subscribers in the app; Carbon is not.
Counterpoints and risks
- Ten tests is still ten tests. Hands-on runs are revealing but not statistically rigorous. Different prompts could flip several results.
- Verbosity is a real cost. Long answers take longer to read and can hide errors.
- No citations in research mode is a meaningful weakness for anyone fact-checking.
- Medical results are illustrative only. A 1.5/6 on tumor identification is a curiosity, not a capability you should rely on.
- Benchmark sources vary. Some numbers come from Google's own model card; others from third parties at different settings. Treat cross-model comparisons as approximate.
FAQ
Is Gemini 4 Neon better than GPT-6.2 Astra? It depends on the task. In these tests Neon won on the piano duet, song, brain scans and speed, and lost on the Unreal Engine game and deep-research style. On the main intelligence index they're close, with Astra slightly ahead.
Why is Neon so much faster? Google designed it as an efficient mid-tier model. In practice it finished several tasks in a fraction of the time rivals needed, most strikingly the Blender villa (about 21 minutes versus over an hour).
Does Neon hallucinate less? On Artificial Analysis's Omniscience test, yes. Its hallucination rate is around 14%, compared with roughly 50% to 59% for top Claude and GPT models.
Where can I use it? In the Gemini app for Pro and Ultra subscribers, in Google AI Studio and the API on paid plans, and in Google's Antigravity agent tool.
The bottom line
Gemini 4 Neon isn't the smartest model you can buy. On raw knowledge, Claude and OpenAI's best still lead. But in hands-on work it kept doing something rivals didn't: finishing hard jobs quickly, cheaply and without making things up.
It played two pianos at once, rebuilt a real house in 21 minutes and copied a video's effects nearly perfectly. It also botched a sword fight and missed a frog sitting in plain sight.
That's a very human-shaped profile: fast, efficient, honest about its blind spots, and still not good at everything. For a model that's supposed to be the middle option, that's a strong place to be.
Sources
- Google: Introducing Gemini 4 Neon
- Google DeepMind: Gemini 4 Neon model card
- Google Antigravity blog: Gemini 4 Neon in Google Antigravity
- 9to5Google: Gemini 4 Neon announcement
- BenchLeader: Gemini 4 Neon independent results
- Gemini 4 Neon vs Carbon, Claude Opus 5.5 and GPT-6.2 Sol (Artificial Analysis data)
- ofox: Gemini 4 hallucination rates vs Claude and GPT





