In partnership with

What's Actually Happening

Alibaba just put its biggest model on the open market. Qwen4-Max left preview today with a 1-million-token context, text, image and video input, and a list price of $1.50 per million input tokens and $4.50 per million output, a fraction of what the Western frontier labs charge.

That makes it one of the cheapest near-frontier models you can call through an OpenAI- or Anthropic-compatible API tonight, and Alibaba is throwing in a million free tokens to try it.

Also tonight: Meituan opens the weights of its 1.6-trillion-parameter LongCat-2.5, Gemini 4 Carbon takes the MathArena crown at a steep price, AI labs hire spy-catchers, and four tools worth trying this week.

Alibaba moved Qwen4-Max to general release today under the model ID qwen4-max, ending the Token Plan-only preview that opened November 11. It reads text, images and video, writes text, and holds up to 1 million tokens of context. Alibaba Cloud lists it at $1.50 per million input tokens and $4.50 per million output in its Singapore region, with a 1-million-token free quota good for 90 days.

Under the hood, DataCamp's breakdown of Alibaba's launch materials puts it at a 2.4-trillion-parameter mixture-of-experts backbone with 71 billion parameters active per token. It speaks four API dialects (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and Alibaba's DashScope), so most existing agent code needs only a new base URL and key. Thinking, parallel tool calls, web search, structured JSON output and prompt caching are built in, and output tops out at 131,072 tokens with thinking off. It's also live inside Alibaba's QwenWork, Qoder and QwenStudio apps.

On Alibaba's own chart, Qwen4-Max edges Claude Opus 5.5 on Agents' Last Exam (46.7 vs 43.4) and DeepSWE (75.7 vs 74.2), but trails it clearly on Terminal-Bench 4.0 (48.5 vs 59.6) and Humanity's Last Exam (51.7 vs 59.8). The independent tracker Artificial Analysis ranks it ninth on its Intelligence Index, per DataCamp.

Our take: this isn't the smartest model you can buy, and Alibaba doesn't pretend otherwise. It's a near-frontier model at mid-tier prices, which is exactly the slot that matters for high-volume agent work where Opus-class bills hurt. Three caveats: the weights are closed (Alibaba says smaller Qwen4 models will be open-sourced later), most of the benchmarks are Alibaba's own, and your prompts and files run through Alibaba Cloud, so check where your data is allowed to go before you ship it to customers.

💸 Why It Matters: The Price Floor Just Moved

Output tokens are where agent budgets go to die, because reasoning models bill their thinking as output. At $4.50 per million, Qwen4-Max's output price is under a quarter of Claude Opus 5.5's $20 and less than a tenth of GPT-6 Astra's $50. Gemini 4 Carbon looks close at $10 today, but that's an introductory price that ends November 30, when Google moves it to $20.

Alibaba's own pitch is cost per finished job, not per token: on its in-house Work360 benchmark, it says average cost per task fell from $4.43 with Qwen3.8-Max to $1.67, a 62% drop, partly because the new model uses fewer tokens for the same work. That comparison is against Alibaba's older model, not rivals, and no outside red team has tested Qwen4-Max yet. But the direction is clear: Chinese labs are no longer just undercutting with open weights. They're now pricing a closed flagship to pull agent workloads off Western APIs.

A Word From Today's Sponsor

1,000+ Proven ChatGPT Prompts That Help You Work 10X Faster

ChatGPT is insanely powerful.

But most people waste 90% of its potential by using it like Google.

These 1,000+ proven ChatGPT prompts fix that and help you work 10X faster.

Sign up for Superhuman AI and get:

  • 1,000+ ready-to-use prompts to solve problems in minutes instead of hours—tested & used by 1M+ professionals

  • Superhuman AI newsletter (3 min daily) so you keep learning new AI tools & tutorials to stay ahead in your career—the prompts are just the beginning

Let’s Talk MODELS! 🤖

Meituan released the weights for LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model with about 48 billion parameters active per token, a week after unveiling it. It's built for coding agents, takes text, images and video, and handles up to 1 million input tokens and 128,000 output tokens through the API. The license is MIT, so commercial use, modification and redistribution are all allowed.

On Meituan's own numbers, it scores 88.4 on Terminal Bench v2.1, within a point of Claude Opus 5 (89.1) and GPT-5.6 Sol (88.8), and 64.6 on DeepSWE, up from 9.7 for LongCat-2.0. The catch is hardware: the sample self-hosting setup uses 194 Huawei Ascend chips, and the guide offers no instructions for Nvidia GPUs. A 4-bit build is also out, which Meituan says processes prompts about 43% faster than its 8-bit version.

Our take: an MIT-licensed 1.6T model is a real gift to anyone who wants to fine-tune or host a big agent model without license strings. But the rivals in Meituan's table are a generation old, the scores are self-reported, and the deployment path assumes Chinese accelerators, so most readers will meet it through the chat app or API rather than their own racks.

The Math Crown Changed Hands

🧮 Gemini 4 Carbon Tops MathArena

Google's Gemini 4 Carbon, released to the public on November 10, took first place on MathArena, the math-competition benchmark, knocking OpenAI's GPT-6.1 Sol off the top. It's the first non-OpenAI model in a while to hold the No. 1 spot, with Anthropic's Claude Fable 5.2 in second.

The win came with a big bill. At the reasoning settings MathArena used, Carbon cost roughly 15 times as much per evaluation as GPT-6.1 Sol, and its most common failure was blowing through the time limit; in one case it ran a script for more than 11 hours without checking on it. It almost never gives up, which is great for hard proofs and dangerous for your API budget.

Our take: if you need the best math reasoning money can buy, Carbon is now the pick, but set hard timeouts and spending caps before you let it loose. And budget at the post-November 30 rate of $4 in and $20 out per million tokens, not today's introductory $2 and $10.

Top 5 In AI Research 🔬

The other stories worth your attention:

AI labs are hiring counterintelligence staff at a record pace, with Anthropic's count more than doubling this year and OpenAI and Nvidia hiring too, as espionage fears spread through the industry. See what these roles now pay. (Fortune)

The FAA is taking its $875 million AI air traffic tool nationwide, expanding SMART from Washington, D.C. airspace to all 22 centers, with controllers keeping the final say. See how much time and fuel it saved per rerouted flight. (Gizmodo)

Shopify switched on a rulebook for AI shopping agents in every store, as Sierra released Poppy 1.0, a draft protocol that tells personal agents what a business lets them do; Meta's Muse checks for it first on every site. See which airline already lets agents hunt for upgrades. (Sierra)

Uber detailed how it polices its own AI agents, with a framework that registers every agent, checks whether each action fits the request before it runs, and logs everything for replay. See where its checkpoints sit in an agent's life. (InfoQ)

Siri AI reached older Apple TV 4K boxes with tvOS 27.2.1, adding personal-context suggestions and task handling to the 2nd- and 3rd-generation models. See which iPhone you need paired to use it. (MacRumors)

🛠️ Tools That Are Hot Right Now!

  • 🩺 Apple Health AI Study - ask health questions, rate AI answers built with OpenEvidence, and help shape how Apple handles health AI. Free for US adults over 21 on an iPhone 15 or later; see what Apple says it won't do with your data.

  • 🤖 Rakazo - run AI teammates on your own machine that work across Gmail, Slack, GitHub and Notion, with approval rules and your own models. Free to self-host, cloud from $15/month; see how it keeps credentials local.

  • 👥 Frevana Multiplayer - share prospecting and content agents with your team, so a workflow set up once gets reused by everyone. Free options available; see how humans take over a running agent.

  • 🚚 Qwen Live Delivery - steer long Codex CLI tasks from a browser workbench: review versions, add requirements and set a time budget mid-run. Free and open source (Apache-2.0); see how it splits a big task into versions.

What's The Recap?

Alibaba took Qwen4-Max out of preview, pricing a near-frontier, 1M-context model at $1.50 in and $4.50 out per million tokens, with a million free tokens to start. Meituan opened the MIT-licensed weights for its 1.6T LongCat-2.5, and Gemini 4 Carbon took the MathArena crown at roughly 15 times the cost of GPT-6.1 Sol. The thread: the very top keeps getting pricier, the tier just below it keeps getting cheaper, and that second tier is where most agent work actually runs.

Login or Subscribe to participate

Stay building. 🤖

Our newest video breaks down the latest AI releases and what they mean for builders.

📬 Read by 14K AI readers. Know someone who'd like it? Send them here to subscribe.

Recommended for you

View all
caret-right