New AI models this week: GPT-6.2 Astra and Sol, Gemini 4 Neon, Qwen4 Max
OpenAI's GPT-6.2 Astra and Sol, Google's Gemini 4 Neon and Alibaba's Qwen4 Max led the week of Nov 15-22, with prices from $1.43 to $8.00 per 1M input tokens.

OpenAI’s GPT-6.2 Astra and Sol led the week of Nov 15 to Nov 22, 2026, followed a day later by Google’s Gemini 4 Neon. Alibaba’s Qwen4 Max arrived at the start of the week with a 1M-token context window and the lowest API price of the group. Open-weight releases on Hugging Face rounded out a busy seven days.
GPT-6.2 Astra and Sol: OpenAI’s new pair for hard projects and everyday work

OpenAI announced GPT-6.2 Astra and GPT-6.2 Sol on Nov 17, with ChatGPT and API access from the first day. OpenAI describes Astra as its new frontier model for the hardest projects, aimed at complex professional work, coding and computer use. Sol is the efficient, versatile model for everyday work and agents at scale. Both take text and image input, return text, and offer a 1.05M-token context window with up to 128K tokens of output.
Pricing per 1M tokens is $8 input and $40 output for Astra, and $2 input and $10 output for Sol. OpenAI says Astra’s prices are 20% lower than GPT-6 Astra’s. On benchmarks, OpenAI reports that Astra scored 1,318 Elo on Artificial Analysis’ AA-3D blind vote, against 1,271 for Claude Fable 5.2 and 1,259 for Sol. It also says Sol comes within 0.5 points of Astra on Cognition’s FrontierCode 2.0 at about one-fifth the cost. These are the maker’s own figures. OpenRouter’s catalogue also lists “Pro” entries for both models at the same prices, but OpenAI’s announcement does not describe a separate Pro model, so they are not treated as distinct releases here. Details are in OpenAI’s announcement.
Gemini 4 Neon: Google’s mid-tier frontier model for software and knowledge work

Google introduced Gemini 4 Neon on Nov 18. It sits between Gemini 4 Carbon, which Google positions for the hardest long-horizon work, and Gemini 4 Hydrogen, built for fast everyday execution. Google says Neon delivers frontier-level intelligence with fast task completion across software engineering, knowledge work and science, and that its cost per task at the default high thinking level is 32% lower than Carbon’s.
The model has a 1M-token context window and accepts text, image, audio, video and PDF input. Google lists $3 per 1M input tokens, $0.15 for cached input and $15 per 1M output tokens. It is rolling out in the paid Gemini API tier in Google AI Studio, Gemini for business, Google Antigravity and the Gemini app for Pro and Ultra subscribers. Google reports an Artificial Analysis Intelligence Index score of 61 at its highest setting, ranked third overall, and first place on Zapier’s AutomationBench at 56.3%. Those are Google’s claims. Our report on the Devin and FrontierCode results covers how it compares with Claude Opus 5.5 on cost. See Google’s announcement.
Qwen4 Max: Alibaba’s flagship multimodal model with a 1M-token window

Alibaba’s Qwen4 Max is a flagship multimodal model that, per Alibaba Cloud’s documentation, combines text generation, image and video understanding and deep thinking. Alibaba positions it for programming, office work, front-end design and complex agent tasks. It accepts text, image and video input and outputs text, with a 1M-token context window and up to 262,144 tokens of output in thinking mode.
The tracker shows $1.43 per 1M input tokens and $4.28 per 1M output tokens, the cheapest of the week’s major models. Alibaba’s own page lists slightly different rates by region, from $1.413 to $1.50 for input. The documentation states no benchmark results, and the model is not offered for fine-tuning. See Alibaba Cloud’s Qwen4 Max page.
Open-weight releases worth a look on Hugging Face
Tencent published Hy4-flash on Hugging Face this week, following its large base model. Our coverage of Tencent’s 770B Hy4-base reports that it beat DeepSeek’s V4 Pro base in 18 of 21 tests. Thinking Machines’ Inkling-2-Mini also drew attention, and we explained how it runs three times over its middle layers to fit a laptop.
Community work built on Qwen3.8-27B continued as well. Our report on a small LoRA adapter that lifts Qwen3.8-27B from 39% to 55% on Terminal-Bench covers one such release. Other new uploads people picked up include Apple’s FastFrames-L, Meituan’s LongCat-Video-Avatar-1.6, SenseNova-Loop-9B from SenseTime’s sensenova organization and Inworld’s voice-isolator. We have not tested these, and we list them only as releases.
Other launches our watcher saw
Alibaba’s Qwen team opened a GitHub repository called Qwen-Live-Delivery-Plugin. OpenAI published auth-javascript, a Sign in with ChatGPT SDK for JavaScript, TypeScript and React.
Not counted this week
The tracker also lists Mistral AI’s Pixtral Large 2502 with a Nov 20 date. The “2502” in the name points to an earlier snapshot rather than a new model, so it is left out of the roundup. Current prices for all tracked models are on our AI API pricing tracker.
The week in numbers
- Models covered: 4 (GPT-6.2 Astra, GPT-6.2 Sol, Gemini 4 Neon, Qwen4 Max), from 3 makers.
- Cheapest per 1M input tokens: Qwen4 Max at $1.43 (GPT-6.2 Sol is next at $2.00).
- Most expensive per 1M input tokens: GPT-6.2 Astra at $8.00 ($40.00 for output).
- Context windows: about 1M tokens for all four.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.



