
Daily AI Brief, Nov 16: Qwen4-Max Gets a Meter and Tencent's Hy4 flash Lands at $0.07
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 96 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 130 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 150 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 120 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 219 tok/s
Two Chinese labs put a cheap multimodal model with a million-token window on sale the same day, and the spread between them is the story: Qwen4-Max, Alibaba's 2.4-trillion-parameter sparse mixture-of-experts, meters at $1.50 per million input tokens and $4.50 output, while Tencent Hy4 flash — a 224-billion-parameter model with 12 billion active per token — lists at $0.070 in and $0.281 out with cached input at $0.003. That is a 21-fold gap on input between two models that both claim native vision, both accept a million tokens of context, and both went live on 16 November 2026. Neither number is a discount coupon or a launch-week teaser; both are the standing rate cards, with promotional layers sitting on top.
They are different products aimed at different buyers, which is why the gap is real rather than a pricing error. Qwen4-Max is a flagship competing for the work a frontier model does, and it is priced 25% under Alibaba's own previous flagship. Hy4 flash is a distilled workhorse built to be boringly cheap on office-shaped tasks, and Tencent's research post says so in as many words: use it for documents and slides, keep the flagship for long software engineering. What follows is what each company published, what it declined to publish, and which parts of either release you can actually reach from an API today.
Item one — Qwen4-Max stops being a preview
Alibaba moved Qwen4-Max from a token-plan preview to metered general availability on 16 November, and the identifier changed with the billing model: qwen4-max is the live string, while qwen4-max-preview and qwen4.0-max-preview are documented for retirement at 24:00 on 25 November 2026 UTC+8. Anyone with either preview string in a config has nine days to change it.
• Architecture — 2.4 trillion parameters total, 71 billion activated per token, plus 328 billion N-gram embedding parameters that are looked up rather than multiplied. Three Gated DeltaNet layers compress history into fixed state for each Qwen Sparse Attention layer.
• Context and output — a 1,000,000-token window. The shared thinking-plus-answer budget is 262,144 tokens, down from 393,216 on the preview card; non-thinking output caps at 131,072 with a 991,808-token input ceiling.
• Price, Singapore — $1.50 input and $4.50 output per million tokens, flat across the whole window, with cache reads at $0.075. The other five regions list $1.413 and $4.24.
• Price, the previous generation — Qwen3.8-Max is $2.00 and $6.00 in Singapore, so the headline is a straight 25% cut on both sides of the meter. In the Global regions the cut is 14.4%, because the older model's non-Singapore card was already discounted.
• Regions and surfaces — Beijing, Singapore, Frankfurt, US Virginia, Tokyo and Hong Kong; OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope; text, image and video in, text out. Batch inference exists in Beijing only.
The vendor benchmarks are Alibaba's own and nothing outside the company has reproduced them. On Work360, its internal taxonomy of 26 Tier-1 and 721 Tier-3 domains, Qwen4-Max scores 52.69 against 47.74 for Qwen3.8-Max, with roughly 24% fewer tokens at xhigh effort and 35% fewer at high. The average cost per task on that suite falls from $4.43 to $1.67, a 62% reduction — a figure that multiplies an unverified price cut by an unverified token-efficiency claim, which is worth remembering when it is quoted at you without the multiplication shown. Artificial Analysis had no Qwen4-Max entry when this brief was written.

Item two — Tencent ships the cheap end of the Hy family
Tencent's research post is dated 16 November 2026 and is titled, roughly, the value-for-money model for everyday work. The specification it leads with is narrow and specific, which is more than the Qwen release does with its benchmark suite.
• Size — 224 billion parameters total, of which 64 billion are Engram parameters, with 12 billion active per token.
• Context and modality — 1,048,576 tokens, multimodal input. Tencent does not say how large the vision encoder is or whether video is accepted.
• Price, mainland card per million tokens — ¥0.5 input, ¥2 output, ¥0.02 on a cache hit. The same card in dollars is $0.070, $0.281 and $0.003.
• Price against its own flagship — Tencent HY4 lists ¥6 input, ¥18 output and ¥0.3 cached, so the flash tier is twelve times cheaper on fresh input, nine times on output and fifteen times on cached input, at the same context window.
• Weights — none. Tencent says it plans to open-source Hy4 flash and to iterate on it frequently. No licence, no date, no repository.
• Independent score — none. Artificial Analysis carries Tencent HY4 at 34.6 on its Intelligence Index as of 5 November and has no entry for Hy4 flash.
Tencent names four levers behind the price. Hybrid attention, LatentMoE and the Engram parameters give a KV cache one twenty-fourth the size of Hy3's and one sixth of Tencent HY4's. Self-distillation from Tencent HY4 followed by reinforcement learning cut end-to-end output length by 50%. A megakernel, four-bit weights and a decoding path called DFly hold decode above 150 tokens per second. And the evaluation was moved off public leaderboards onto WorkIndex, an internal eight-suite set covering slide creation and revision, co-writing, document editing, table repair, audio and video processing, and visual understanding — with the public agent benchmarks (DeepSWE, ProgramBench, Terminal-Bench 4.0, APEX-Agents, AutomationBench, HLE, OSWorld 2.1) published as chart images rather than a table.
The claim that travelled is a blind test: 85 internal WorkBuddy users, 27 task categories, 259 real daily-work tasks, with satisfaction reported as close to Claude Haiku 5.5, the lowest per-task cost in the panel and the shortest task duration. Close is the operative word — the launch summary says the model matches Haiku 5.5, the research page says satisfaction was close to it. The panel is the vendor's own employees, in the vendor's own product, on tasks the vendor selected. That is a reasonable way to decide what to ship internally and weak evidence about anyone else's mileage.
Tencent's per-task cost claim is more careful than most: on WorkIndex, one-thirteenth the cost of Tencent's own Hy4 preview, one-twentieth of Tencent HY4, and one-eighth of DeepSeek-V4.1-Flash-1017. Note the parenthetical the vendor attaches to the last one — maximum effort tier, busy-hour price. DeepSeek's busy-hour rate is twice its off-peak rate, so the one-eighth is measured against the more expensive of DeepSeek's two published cards. That is not dishonest, because a busy-hour deployment is a real deployment, but it is the comparison that flatters Tencent most.
What both releases have in common, and what neither has
Two labs, one day, one shared gap: neither model has been measured by anyone who does not work for the company that built it. Alibaba's suite is a taxonomy of its own construction; Tencent's is an internal benchmark it explicitly chose over public leaderboards. Tencent goes further and publishes its benchmark results as PNG charts, so the numbers cannot be copied, diffed or re-run without reading pixels.
The other shared property is that both are unopen. Qwen4-Max ships as an API only, with the smaller Qwen4 models promised for open-sourcing on the same lineage. Hy4 flash ships as an API only, with a plan to open-source and no date. Both companies have released weights before — Tencent's August Hy4 preview went out under Apache 2.0 with serving recipes on day one — which is why the hedged language in both releases reads as a decision rather than a limitation.
Where they diverge is what the money buys. Qwen4-Max is a general-purpose flagship: the pitch is capability at a lower price than the last flagship, with a token-efficiency argument layered on. Hy4 flash is a specialist: the pitch is cost per completed task on document, slide and agent work, with the vendor conceding a small drop on long coding tasks in its own internal data. A 21-fold input-price gap between two million-token multimodal models is not a market inefficiency. It is two companies pricing two different jobs.
What you can call today
Neither model is on OrcaRouter yet. The catalogue's Tencent routes are tencent/hy4-preview at $0.834 input and $2.501 output per million tokens — the model Hy4 flash was distilled from — and tencent/hy3 at $0.18 and $0.59. The Alibaba route is qwen/qwen3.8-max at $2.00 and $6.00. Qwen4-Max and Hy4 flash are reachable through their vendors' own APIs and, in Tencent's case, a list of its own consumer and enterprise products: TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis. Tencent says third-party platforms will follow and names no date.

The route that matters for this week's decision is the DeepSeek one, because Tencent's own cost comparison picks DeepSeek-V4.1-Flash-1017 as its benchmark: deepseek/deepseek-v4.1-flash is on the catalogue at $0.15 and $0.60 per million tokens with cache reads at $0.003 — DeepSeek's own off-peak list price, passed through with no markup — over a 1,048,576-token context, alongside the 1017 snapshot at the same rate. If you want to test whether a twelfth of the input price produces a tenth of the per-task cost on your traffic, that is the number to measure against, and it is callable now while the two new models are not.
Two promotional layers sit on top of the Tencent card and neither should enter a budget. Tencent Cloud's TokenHub is running 50% off for an unstated period, and WorkBuddy is free until 30 November. Alibaba's Singapore card carries a one-million-token free quota valid for 90 days. All three expire; the list rates are what survive.
Two dates to diary
The first is 25 November, when the Qwen4-Max preview identifiers stop working. If you are on the preview, the migration is a string change plus two checks that will not raise an error when they are wrong: any code that assumed a 393,216-token shared output budget now has 262,144, and anything that relied on Token Plan credits is now metered per token.
The second is 30 November, when WorkBuddy stops being free and the TokenHub promotion may or may not still be running.

Beyond those dates, the thing to watch is not a price. It is the first number from a harness that neither company wrote. Qwen4-Max will get an Artificial Analysis entry within days, simply because Alibaba models always do. Hy4 flash may not, since Tencent's own flagship only reached that index on 5 November, eight days after it shipped. Until one of them is measured outside the building, the honest reading of 16 November is that two very cheap million-token multimodal models went on sale at prices that undercut their own predecessors, and that the evidence for each is a set of charts drawn by the company selling it.
Compared in this article3
Detected from this article · Benchmarks: Artificial Analysis · updated daily
