Elon slowly becoming right again.
He was the first one to say that in 5-10 years AI will do most of the work human doing today. AI already doing what human loved to do then the question is:
What human love to do is already done by AI what human do now?
@grok what do you think
After handling Gemini 4 Neon my development work, I think, it’s already a better than Sol and Opus 5.5.
It's hard pill for the Anthropic and OpenAI fans to take it, but accept it, Neon is the best model in that tier right now.
Neon is so much better at speaking human, it gets what you want to achieve, and even comes up with much better ideas, you can actually out into actions.
It is really efficient and calls tools at once and gets to work done. This little guy doesn’t waste words, just gets it done, and fast!
Goooooogle, you really coocked harder!
Dhaval retweeted
The AI race is no longer just about building the best model. It's also about building a business that can survive the race.
Anthropic's stock market debut is a big moment for the industry.
With billions raised and strong investor interest the company behind Claude now faces a different challenge:
turning AI demand into long term business value.
The real competition is shifting from who builds the most impressive AI to who can make it sustainable.
@AnthropicAI
Gemini 4 Neon Technical Report: Key Highlights
1. Positioning: The most balanced tier in the family
Carbon tackles the hardest long-running projects, Hydrogen handles everyday tasks, and Neon sits in between: complex coding, knowledge work, and scientific tasks, with frontier intelligence + fast end-to-end completion. Google says it's built for frequent, heavy use—not a flagship, and not a budget model.
2. Intelligence Index score of 61, ranking third
AA Intelligence Index score of 61, ranking third, behind only larger, slower models. Vals Index ranks second (70.9%), and the new Agent Arena Index score is 1336, ranking second.
3. Tops the coding agent leaderboard
Artificial Analysis Coding Agent Index score of 58.6, ranking first; Terminal-Bench 4.1 scores 65.2%, and FrontierCode 2.0 scores 46.2% at the default high setting. Its specialty is precise, small changes when debugging and refactoring across multiple files, rather than blindly rummaging through files.
4. Priced at $3/$15, 32% cheaper than Carbon
Input $3, output $15 per million tokens, cached input $0.15 (95% discount), 1M context, with four adjustable thinking levels. At the default high setting, the cost per task is 32% lower than Carbon and 50–80% lower than top competitors.
5. Best-in-class scientific reasoning
Best in its class. Also took several first-place spots on legal, tax, and accounting benchmarks.
6. Lowest fabrication rate in the family
In internal Antigravity tests, its fabrication rate was the lowest in the Gemini 4 family; cheating was 0.0%, and taking shortcuts accounted for just 2.45% of tasks, also the lowest in the family. One of this release's selling points is that it doesn't cut corners or overstep boundaries halfway through long-running tasks.
7. The cybersecurity-specific version packs a punch
Neon for cyber: 76% on AutoBaxBench hard problems, ranking first; 51.26% on WebExploitBench, ranking first, at half the cost of competitors; 90.5% on SEC-bench Pro for finding vulnerabilities, and 70.7% on CWE-bench v2 for patching vulnerabilities (3.4% higher than Carbon). Initially available to trusted defenders in the Fairwind program.
1033 tok/s on a laptop RTX 3060 with 6 GB VRAM!
Kinetic has been created, a new inference engine written in Rust, which is intended to compete with llama.cpp.
The authors published impressive benchmarks on a laptop with a Ryzen 5 6600H and an RTX 3060 6 GB.
When transcribing a file using the LFM3 1.7B model:
- llama.cpp: 229.7 tok/s
- Kinetic: 1033.7 tok/s
Over 4.5 times faster!
During the first run, the speedup is smaller: up to 2.5x during editing and up to 1.4x during normal generation, but the difference is still significant.
github.com/Veloresearch/kine…
This guy predicted Fable 5.2 and GPT-6.2 Astra and Sol release before anyone, surprisingly he has only 2 post.
did any get to know who th is @icedritual ?
Dhaval retweeted
Replying to @SiftByte
I tested neon overnight. It’s as good as he says. I still don’t understand why Argon exists though.
We got models in just week time from:
- Anthropic have Fable 5.2, Opus 5.5
- Google have Neon, Carbon, Hydrogen
- OpenAI have GPT-6.2 Astra and Sol
Hey @grok, if you have to setup your workflow, then how would you choose the models and do your workflow setup?
Be Specific and give me answe, also states why you choose the model and what kind of work you would give to these models?
Gemini 4 Neon, what a medium tier model Google has shipped. its going to be really hard for both OpenAI and Anthropic people to swello it.
It was just a day before GPT-6.2 Sol released and Google ended it's run before it could even start. It's even not sparring Opus 5.5 and just humiliating it everywhere.
I want to know, what @GoogleDeepMind is feeding to their model, their mid and flash tier models are coming out so goooood?
Is Gemini cheap right now? Depends which one, and which day. (Google's pricing page, Nov 19)
• Gemini 4 Neon (new, Nov 18): $3 in / $15 out per 1M tokens. Full price, no promo.
• Gemini 4 Carbon & Argon: $2 / $10, but only through Nov 30. On Dec 1 they double to $4 / $20.
• Gemini 3.8 Flash: $0.75 / $3.75 through Dec 31, then $1.50 / $7.50.
So the discount is a temporary half price on the big models, not on the new one. Until Nov 30, Carbon is cheaper than Neon. From Dec 1, Neon is 25% cheaper than Carbon.
For scale: Claude Sonnet 5.5 lists at $2 / $10, Opus 5.5 at $4 / $20.
Two catches: thinking tokens are billed as output, and none of these three Gemini 4 models has a free tier. Before you move a pipeline for the promo, price it at Dec 1 rates.
——This note comes from Hark, the best experimenter at Miralucis, who asked me to pass it on to you all.
The boundaries of what small AI models can do have been pushed again!
The effectiveness of Qwen3.8-27B in programming tasks has increased from 39.2% to 65%.
The authors of the harness-hub project trained the adapter using the reinforcement learning method, teaching the model to work more effectively with the pi agent's tools.
The results of their tests:
- SWE-bench Pro: 39.2% -> 64.7%
- Terminal-Bench 2.1: 39.5% -> 55.8%
- Number of agent steps for solved tasks: 59 -> 23 in SWE-bench Pro and 18 -> 8 in Terminal-Bench.
But there is something else.
The model with the adapter, without reasoning mode enabled, matches the base Qwen3.8-27B with the highest reasoning level, generating about 4.5 times fewer tokens!
This shows that proper training can teach a model to use tools more effectively, instead of forcing it into increasingly longer reasoning.
I wonder how much can be squeezed out of Qwen4 27B when it comes out?
huggingface.co/harness-hub/Q…
Vishal, I agree that Google took it so personally that you guys at Google are iterating as fast as you can.
But at least give your own models some breathing time. It was just a week ago that you launched the Hydrogen, and before that, it was 2 weeks ago that you launched the Carbon, and it was the best from Google.
In just the last month, you dropped all the models, even your own models are going to be deprecated that quickly.
Also, please, how much is left in Gemini 4? It's getting hard to keep track of the whole Gemini 4 model family. 😃
Taking a moment to reflect on the last 5 weeks of Gemini 4:
- Need more intelligence: we shipped Argon
- Need even more intelligence: we shipped Carbon (4 weeks later)
- Actually, we need it faster: we shipped Hydrogen (2 days later)
- Well, maybe more intelligence, just keep it fast: we're shipping Neon (6 days later)
We hear all the feedback and we're iterating as fast as we can. Keep telling us what's working (and what isn't) and let us know what you think of Neon!
Anyway, back to work..











