This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

DeepSeek's team published "Super-agent Design Exploration: Toward More Autonomous Agent Systems" and open-sourced a runnable prototype on GitHub (dsh-superagent-prototype, MIT). It's not a new product - it's an engineering blueprint answering a very practical question: AI can finish a single task without supervision now, but one person can only follow 3-5 active sessions at once - the multi-task ceiling is shifting from "model capability" to "human attention."
DeepSeek's answer borrows decades of systems engineering:
In a simulated collaboration of ~300 messages and 20 work requests (vendor-described experiment, synthetic, simplified, 5 runs per config, marked "preliminary"): Super-agent median completion 12.4 min, p95 15.7 min; the traditional single-session approach was 20.4 / 29.5 min - nearly twice at p95. Subscriptions saved the Router ~37% token overhead; batching cut per-decision inference from 6.4s to 1.6s. The team also admitted the other side of the ledger: more cached-token reads than single-session.
DeepSeek says an internal instance built on this has been running in daily development - reproducing and fixing issues, code-review responses, and iterating the Harness itself. Take: while others race to make models "a bit smarter," DeepSeek is investing in "system-layer scheduling." If this route pans out, agent products morph from "a chat box" to "a colleague that lives in the background."
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| SenseNova Loop-9B-Preview (SenseTime, open source, Apache 2.0) | 10-30 / 11-1 | Multimodal LLM (recurrent) | Pitched as "deepen compute, don't stack parameters": 32 layers holding 9B, the middle 24 layers run twice on shared weights, one forward pass equals 56 layers (~14.6B inference), with an mHC dual-stream residual to stabilize signal. Vendor-claimed office and agent tasks beat Qwen3.5-27B (WorkBuddy Bench 66.89, BIRD-Dev 71.38) but lose Terminal-Bench 2.1 (40.45 vs Qwen3.6-27B's 47.19); all self-reported, no tech report yet. 262K training context. |
| Pika H3 Flex (Pika, open weights + API) | 10-29 / 11-1 | Audio-video joint generation | A post-trained "iteration specialist" on MiniMax H3: two LoRA adapters (text/first-last-frame FL2VA, multimodal-reference Ref2VA with up to 12 references), distilled to 6 denoising steps (teacher 50), native 480P/768P up to 15s with synced stereo; the API can split audio into dialogue/score/SFX/foley/ambience tracks. Vendor: #8 on DesignArena text-to-video, #1 open video model; "15s 768p in 7s" is a controlled-benchmark claim, unreplicated. |
| PixelUMM (NVIDIA x Waterloo, open source) | 9-29 / 11-1 | Unified multimodal (understand + generate) | Kills the "standard dual-encoder": no ViT, no VAE - images split into 16x16 patches, video into 4-frame spatiotemporal blocks, one linear projection feeds raw pixels into a Qwen3-8B backbone, understanding and generation share attention. Top unified-model video understanding (MVBench 70.53), VBench 83.24, GenEval 0.77 - the paper says "comparable" to baselines, not crushing. Code and weights open. Honest weakness: MMMU 41.67, well behind specialist VLMs; grid artifacts on smooth regions at high CFG. |
Today's keyword is "digging deeper." DeepSeek touches the agent's system layer, SenseTime touches Transformer depth, NVIDIA touches the multimodal encoder - after the parameter race cooled, everyone is betting on the same thing: smarter structure rather than bigger scale, wringing more output out of each unit of compute.
(Data from official announcements and public reports, 2026-11-02; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)
GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer
Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi
Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload
gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi
Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa