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.

On the afternoon of Nov 11, Baidu released the base model Wenxin 5.5 Flash Preview. Three hard numbers first: 280B total parameters, 14B activated per token, 1M context, vendor-claimed 200+ tokens/s.
One clarification that decides how you read this: the Nov 5 release was Wenxin 5.5 Flash Base - a raw pre-trained shell with no post-training. Today's Preview is that "GA": same skeleton (280B/14B), now with the full post-training stack - data, RL, tool use, long-horizon agent execution. Not a re-run of old news, but the model going from "raw" to "shippable."
Three technical additions: a new hybrid sparse attention, multi-router MoE, and ECHO multi-token prediction (predicting several future tokens per forward pass for throughput). Plus a first: an in-house RSI technique lets a system called ERNIE Evolve participate in the model's own inference optimization - AI helping optimize AI.
Price is the center of this shot. Vendor figures: input ¥2 peak / ¥1 off-peak, output ¥8 peak / ¥4 off-peak per M tokens, stacked with a limited-time 50% off - landing around ¥1 in / ¥4 out at peak. At 280B scale, that price crushes the field. Live on Baidu AI Cloud Qianfan.
An eight-benchmark table (all vendor figures, none third-party replicated; starred entries are each vendor's own):
| Benchmark (vendor) | Wenxin 5.5 Flash Preview | Same-table peers |
|---|---|---|
| Terminal-Bench 4.0 | 46.5 | Claude Opus 5.5 63.6, GPT-6.1 Sol 56.1, Hy4 42.4 |
| DeepSWE v1.1 | 71.2 | DeepSeek V4.1 Flash 1017 77.0, GPT-6.1 Sol 75.2, Hy4 74.9 |
| AutomationBench | 61.9 (first) | Claude Opus 5.5 58.0, DeepSeek 57.2 |
| OfficeQA Pro V2 | 67.9 | GPT-6.1 Sol 70.0, GLM-5.3-Flash 64.4 |
Read the table both ways. Forward: it wins office automation and spreadsheets - its home turf, explaining the "deliver tables, docs, PPTs" pitch. Backward: Terminal-Bench 4.0's 46.5 is 17 points below Claude Opus 5.5, and DeepSWE trails DeepSeek and GPT-6.1 - running long tasks in a pure terminal, fixing huge repos, it's not first tier. It wins "sitting at a desk and getting work done," not "wrestling the system in a CLI."
Three demos more vivid than the scores:
Four splashes of cold water:
Timing note: Baidu World 2026 opens tomorrow (Nov 12); today's release pre-empts the event's hardest card, so tomorrow likely shifts from "model" to "how agents enter organizations."
| Model / Tool | Date | Type | Highlights |
|---|---|---|---|
| MiniMax H3-Image (open source) | 11-10/11 | Image generation / open weights | Worth trying today. 10 seconds for a 4096x4096 (16MP, 4K) image, vendor-claimed 5x mainstream speed. A 33.1B dense diffusion Transformer (DiT), 50 layers, 5376 hidden, 56 heads, H3's 3D multimodal positional encoding minus the audio branch. Speed from distillation: 6 steps, no classifier-free guidance (the API 400s on other step counts). Up to 16 reference images, 32K-char prompt cap, 16MP any aspect, 20MP ceiling. License matters: MiniMax community license excludes EU, UK, Korea, US, with written authorization over $1M revenue or managed hosting. "10s" is vendor-tested, unreplicated. |
| Gemini Robotics 2.5 (DeepMind) | 11-10 | Vision-language-action model / partners only | Google pushes robot models to the edge of "usable": 200+ real tasks, vendor-claimed 30-100% higher success and 20-30% faster than baseline. The weight is production: with Apptronik and Mercedes, part-sorting on live auto lines in Kecskem\u00e9t and Sindelfingen, 96% success on the Apollo 2 humanoid (a joint ER + 2.5 result). Also open-sources ASIMOV-Jailbreak; Gemini 4 Argon hits 99.5% jailbreak resistance (vs ER 2's 82%). Cloud inference, partners only. |
| Baichuan M4 + "Baixiaoyi" | 11-10/11 | Medical LLM + home AI doctor | Baichuan pushes medical models to home scenarios: medical model M4 plus the family-facing "Baixiaoyi" for health Q&A and light consultation. Background: stricter prescription rules and online-pharmacy regulation - licensed vertical models become the main landing channel. Very few details public; record as "released," no capability judgment. |
| Lenovo "Wanquan hybrid token factory" | 11-10 | Compute / token production | Not a model, but worth noting. A three-layer structure with model intelligent routing - complex tasks to big models, simple interactions to small ones. A new sell: not cards, not hours, but "deliverable tokens." |
| Claude office suite (Anthropic) | 11-10 | Product / paid plans | Claude steps from "generating content" into "entering workflows": Claude Sheets and Claude Whiteboards in beta, Claude Dashboards GA as Claude Data, plus interactive connectors, all plans. Whiteboards turn a fuzzy plan into a timeline with owners and dates; meetings are categorized and action items drafted after. Also a Claude Meetings that listens and drafts plans live. |
The clearest line today: Chinese vendors compete on "how many tokens a yuan buys," American companies on "how much we're worth." One side squeezes 280B to a yuan per million tokens and 4K images to ¥0.23; the other lifts its valuation from $1.4T toward $1.5T in the roadshow. Both rest on one judgment - model capability is homogenizing fast, so the real divide becomes unit cost and distribution.
(Data from official announcements and public reports, 2026-11-11; 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