---
title: "Gemini 4 Neon takes the middle seat, Microsoft prices coding low, and Tencent opens Hy4 flash"
description: "Gemini 4 Neon lands at $3/$15 with a 1M context. Plus Microsoft's MAI Code 2 and Decision 1.1, Tencent's Hy4 flash weights and DaisyRose-27B."
canonical_url: https://www.oguzhan.co/gemini-4-neon-mai-code-2-hy4-flash-daisyrose/
author: "Oğuzhan Koçaklı"
author_url: https://www.oguzhan.co/about/
date_published: 2026-11-19T06:50:42+00:00
date_modified: 2026-11-19T06:50:52+00:00
language: en
translations:
  tr: https://www.oguzhan.co/tr/gemini-4-neon-mai-code-2-hy4-flash-ai-gundemi/
---

# Gemini 4 Neon takes the middle seat, Microsoft prices coding low, and Tencent opens Hy4 flash

Google released Gemini 4 Neon on Wednesday and parked it in the middle of its lineup. Microsoft, Tencent and a small lab called RadixArk filled the rest of 16 to 18 November. Nobody is selling the smartest model this week. Everybody is selling smart enough, for less.

I read the announcements, cards and pricing pages. I ran none of these models.

## 🧪 Where does Gemini 4 Neon fit between Carbon and Hydrogen?

Koray Kavukcuoglu announced Neon on 18 November in [Google’s launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/). It sits between Gemini 4 Carbon (“your hardest, long-horizon projects”) and Gemini 4 Hydrogen (“fast everyday execution”), and the [model card](https://deepmind.google/models/model-cards/gemini-4-neon/) calls it “adapted from Gemini 4 Carbon”. Background: [Carbon going GA](https://www.oguzhan.co/gemini-4-carbon-ga-muse-ember-pan-2-praxis-1/) and Hydrogen in [the 9 to 15 November weekly](https://www.oguzhan.co/ai-weekly-9-15-nov-2026-founder-left-agents-wallet/).

[Google’s pricing page](https://aistudio.google.com/docs/pricing) lists $3 per 1M input tokens and $15 per 1M output, thinking tokens included, paid tier only. Context is 1,048,576 tokens, and the [Cloud docs](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-neon) cap synchronous output at 64K. In [Google Antigravity](https://antigravity.google/blog/gemini-4-neon-in-google-antigravity/), an “Orchestrator” lets Neon plan while parallel Hydrogen subagents work.

The benchmarks are Google’s. Google says Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh, #3 overall, and cites #1 on the AA Coding Agent Index at 58.6. Its model card table is refreshingly mixed. Claude Opus 5.5 is slightly ahead on Terminal-Bench 4.1 (65.8% vs 65.2%) and FrontierCode 2.0 (47.3% vs 46.2%). Neon runs away with GraphWalks long context, 95.2% vs 66.8%. GPT-6.2 Sol trails Neon on every row of that table.

The [Artificial Analysis changelog](https://artificialanalysis.ai/changelog) confirms 44, 55, 60 and 61 from low to xhigh and describes “a small Intelligence gain over Gemini 4 Carbon at a ~17% lower Cost per Task”. Google’s post says 32% lower. Same model, two numbers. I would budget with the smaller one.

Now the fine print I worked out from the pricing page. Carbon costs $2 / $10 through 30 November and $4 / $20 from 1 December. So this month Neon costs more per token than Carbon, and from December it costs less. Fewer tokens per task can still tip the bill toward Neon.

If you run Carbon at the promo price, test Neon at high on your own tasks before 30 November.

![Three glass tubes glowing at different brightness with blank price tags on a dark rack](https://oguzhanco.s3.eu-west-3.amazonaws.com/wp-content/uploads/2026/11/19065038/gemini-4-neon-price-tubes.webp)

*Neon sits between Carbon and Hydrogen on price and capability. Illustration: Higgsfield.*

## 💸 Is Microsoft’s MAI Code 2 really 5% of the price?

At Ignite on 17 November, Microsoft AI released “seven new world class models”, its words. [Microsoft’s announcement](https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/) pitches MAI Code 2 for multi-file work and repo-wide reasoning, and Flash for routine fixes. Both are “explicitly not yet tuned” for very long-horizon agentic work, which I like hearing from a vendor.

[GitHub’s Copilot pricing page](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing) lists MAI Code 2 at $0.50 input and $2.50 output per 1M tokens, a promotion that runs until 17 May 2027. Then it doubles.

Microsoft’s headline: MAI Code 2 gets “96% of Claude Opus 5.5’s performance at just 5% of the cost”. Flash is “comparable” to GPT-6 Luna and Claude Haiku 5.5, the pair from [my price floor post](https://www.oguzhan.co/claude-haiku-5-5-gpt-6-luna-price-floor/), at 65% and 20% of their cost. The footnotes matter. All models ran at max reasoning effort, three of the four benchmarks ran on MSBench, Microsoft’s internal eval platform, and MAI Code 2’s cost is “shown at its 50% promotional price”. Company tests, company platform, promo math.

The [model card](https://microsoft.ai/pdf/MAI-Code-2-And-MAI-Code-2-Flash-Model-Card.pdf) lists a 1,050,000-token context. If you live in Copilot, the cheapest serious coding model in the picker is now Microsoft’s own. Check that May 2027 date before you budget.

## 🎯 Why does Microsoft compare its new model to Jev?

Microsoft-Decision-1.1 went GA on 17 November in Microsoft Foundry, on [OpenRouter](https://openrouter.ai/microsoft/microsoft-decision-1.1) and on Vercel AI Gateway, per [Microsoft’s post](https://commandline.microsoft.com/microsoft-decision-model-ga/). It does not write text. It returns calibrated probabilities for fixed answer options, for routing, classification and verification. Foundry input costs $0.042 per 1M tokens. Output is free.

Microsoft says it scores 68 on the Artificial Analysis Decision Index v2.1, “ahead of Jev 1.14, Google’s Gemini Decision 1, and OpenAI’s GPT-6 Luna (Decisions)”. I did not check that index page, so it stays Microsoft’s number.

Monday brought [Perplexity’s Decider](https://www.oguzhan.co/perplexity-decider-v1-3-deepseek-api-tools-tencent-hy-realtime/), and now Microsoft names Jev as the bar. When I compared [Clef and Jev](https://www.oguzhan.co/clef-vs-jev-cloudflare-decision-models/), this was a niche. Free output is what makes a check before every tool call affordable.

## 📊 Which of these can you actually download?

Prices are per 1M tokens in USD, from each vendor’s page.

| Model | Company | Date | Open weights? | Price per 1M tokens (input / output) | Context |
| --- | --- | --- | --- | --- | --- |
| Gemini 4 Neon | Google | 18 Nov | No | $3 / $15 | 1M |
| MAI Code 2 | Microsoft | 17 Nov | No | $0.50 / $2.50 promo until 17 May 2027, then $1 / $5 (GitHub Copilot) | 1.05M |
| MAI Code 2 Flash | Microsoft | 17 Nov | No | $0.05 / $0.25 promo, then $0.10 / $0.50 | 1.05M |
| Microsoft-Decision-1.1 | Microsoft | 17 Nov GA | No (built on Gemma 4 26B A4B) | $0.042 input, output free (Foundry Global Standard) | 33K (OpenRouter) |
| Hy4 flash | Tencent | 16 Nov, weights 17 Nov | Yes, Tencent Hy License 1.0 | about $0.070 / $0.281 (50% off until 30 Nov) | 1M |
| DaisyRose-27B | RadixArk | 18 Nov | Yes, Apache 2.0 | self-host only | 262K |

## 📦 What can you do with Tencent’s Hy4 flash weights?

Tencent announced Hy4 flash on 16 November and posted the weights a day later, after promising them “in the coming days” back in [the 2 to 8 November weekly](https://www.oguzhan.co/ai-weekly-2-8-nov-2026-cheaper-tokens-pricier-risks/).

[Tencent’s press release](https://www.tencent.com/tencent-launches-hy4-flash-to-power-high-speed-cost-efficient-ai-deployment-for-everyday-workflows/) prices the API at about $0.070 per 1M input tokens and $0.281 output. The [model card](https://huggingface.co/tencent/Hy4-flash) says about 220B total parameters (the press release says 224B), 12B active per token and a 1M context.

Tencent’s quality numbers are internal: on its own WorkIndex benchmark, Hy4 flash scores 42.1 against 39.7 for Hy4 preview at 1/13 of the cost per task.

I read the [LICENSE file](https://huggingface.co/tencent/Hy4-flash/blob/main/LICENSE) itself. The “Tencent Hy License, Version 1.0” is derived from Apache 2.0 but “not the Apache License”. Under Section 10, if you offer it to third parties as a service, an API for example, you need a separate written license from Tencent once worldwide revenue passes US$50M in any 12 months, or before it goes into a product with more than 100M monthly active users.

Running it yourself is harder. The [AngelSlim GGUF builds](https://huggingface.co/AngelSlim/Hy4-flash-GGUF) are 125.73 to 156.79 GiB, and llama.cpp support sits in [PR #32517](https://github.com/ggml-org/llama.cpp/pull/32517), not merged as of today. The README lists speeds for 128GB machines without explaining how files that size fit. I own a DGX Spark, have not run this, and want to see someone reproduce those numbers. Read the license before you host it for anyone else.

![Huge cargo crate with open padlock lowered onto a small home workstation](https://oguzhanco.s3.eu-west-3.amazonaws.com/wp-content/uploads/2026/11/19065040/hy4-flash-crate-home-workstation.webp)

*Hy4 flash weights are open, but the home-machine story is still unfinished. Illustration: Higgsfield.*

## 🌼 What did 1,200 tasks do to a 27B model?

RadixArk released its first model on 18 November. Per [RadixArk’s blog](https://www.radixark.com/blog/daisyrose-27b), DaisyRose-27B is Qwen3.8-27B trained with reinforcement learning only, no supervised fine-tuning, no distillation. It is Apache 2.0 on [Hugging Face](https://huggingface.co/RadixArk/DaisyRose-27B), with the same tokenizer and chat template as the base, so it drops in.

The recipe is the story. About 1,200 long-horizon tasks, each built to trigger a mistake the base model makes, with up to 1.5 hours and 100+ tool calls per attempt. Training took 18 days on 144 H200 GPUs. Tasks that could be passed without really solving them were fixed or removed.

On RadixArk’s own harness, Terminal-Bench 4.1 pass@1 goes from 5.6% to 12.0%. The same table shows GPT-6 Luna at 15.2 from a different harness. Better than its base, not better than everything. I have not run it.

## FAQ

### Is Gemini 4 Neon cheaper than Carbon?

Not per token until 30 November. From 1 December, when Carbon moves to $4 / $20, yes.

### Can I run Hy4 flash at home?

Not with stock llama.cpp yet. The GGUF builds are 125 to 157 GiB and PR #32517 is not merged.

Next on my list: Neon’s Safety & Security Report, Carbon’s price jump on 1 December, MAI Code 2 on OpenRouter, that llama.cpp PR getting merged, someone reproducing the 128GB Hy4 flash numbers, and DaisyRose on independent leaderboards.

## Sources

- [https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/)
- [https://deepmind.google/models/model-cards/gemini-4-neon/](https://deepmind.google/models/model-cards/gemini-4-neon/)
- [https://aistudio.google.com/docs/pricing](https://aistudio.google.com/docs/pricing)
- [https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-neon](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-neon)
- [https://antigravity.google/blog/gemini-4-neon-in-google-antigravity/](https://antigravity.google/blog/gemini-4-neon-in-google-antigravity/)
- [https://artificialanalysis.ai/changelog](https://artificialanalysis.ai/changelog)
- [https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/](https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/)
- [https://microsoft.ai/news/microsoft-ignite-2026/](https://microsoft.ai/news/microsoft-ignite-2026/)
- [https://microsoft.ai/pdf/MAI-Code-2-And-MAI-Code-2-Flash-Model-Card.pdf](https://microsoft.ai/pdf/MAI-Code-2-And-MAI-Code-2-Flash-Model-Card.pdf)
- [https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing)
- [https://commandline.microsoft.com/microsoft-decision-model-ga/](https://commandline.microsoft.com/microsoft-decision-model-ga/)
- [https://openrouter.ai/microsoft/microsoft-decision-1.1](https://openrouter.ai/microsoft/microsoft-decision-1.1)
- [https://www.tencent.com/tencent-launches-hy4-flash-to-power-high-speed-cost-efficient-ai-deployment-for-everyday-workflows/](https://www.tencent.com/tencent-launches-hy4-flash-to-power-high-speed-cost-efficient-ai-deployment-for-everyday-workflows/)
- [https://huggingface.co/tencent/Hy4-flash](https://huggingface.co/tencent/Hy4-flash)
- [https://huggingface.co/tencent/Hy4-flash-FP8](https://huggingface.co/tencent/Hy4-flash-FP8)
- [https://huggingface.co/tencent/Hy4-flash/blob/main/LICENSE](https://huggingface.co/tencent/Hy4-flash/blob/main/LICENSE)
- [https://github.com/tencent-hunyuan/hy4-flash](https://github.com/tencent-hunyuan/hy4-flash)
- [https://huggingface.co/AngelSlim/Hy4-flash-GGUF](https://huggingface.co/AngelSlim/Hy4-flash-GGUF)
- [https://github.com/ggml-org/llama.cpp/pull/32517](https://github.com/ggml-org/llama.cpp/pull/32517)
- [https://www.radixark.com/blog/daisyrose-27b](https://www.radixark.com/blog/daisyrose-27b)
- [https://huggingface.co/RadixArk/DaisyRose-27B](https://huggingface.co/RadixArk/DaisyRose-27B)
