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.

OpenAI released a "mid-cup" model, GPT-6.1 Terra, but the hook isn't topping leaderboards - it's the bill. The pitch is "cheap labor that operates a computer like an office worker" - filling forms, entering invoices, reading diagrams, switching between apps, thousands of high-frequency jobs. That kind of work only pays if per-task cost drops to pennies, and Terra is built for exactly that.
The timing is telling - the same day Anthropic filed its IPO, OpenAI cut prices and shipped. Since 2025, the closed-source camp squeezed by open source has shifted from "upgrade and incidentally lower price" to "cut price outright to grab volume." The practical meaning for developers: "AI clicking your mouse for you" finally makes sense on the bill - for batch forms and data entry, try Terra on a week of real workload before paying for a pricier flagship.
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| Genie 4 (Google DeepMind) | 10-27/28 | World model (research preview) | One sentence generates an explorable virtual world you can walk through with a keyboard; change weather, swap cameras, add characters in natural language at runtime; ~2 minutes of consistent scene, twice Genie 3; the series' first synced audio-video (ambience, music, responsive dialogue). Realtime driven by Gemini 3.8 Flash with Gemini Omni "directing." Open only to trusted testers, rolling to Google AI Ultra subscribers in weeks. Note: Google admits long runs drift in detail and complex physics/in-scene text remain unsolved - demos are all official; no params or scores published. |
| Clover-35B-A3B-Preview (DiffuSpace) | 10-28 | Diffusion LLM (open source) | Domestic diffusion route (global parallel generation, not left-to-right): 35B total / 3B active MoE, 128K context; vendor-measured 680 tokens/s, claimed 5-10x faster than same-size autoregressive; self-reported first across six agent-coding benchmarks (SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7); third-party WindTunnel measured 1.52s median over 49 tasks. Weights on HF/GitHub/ModelScope, SGLang and llama.cpp support; a decision variant is teased. All vendor- or small-sample; no independent big board yet. The company closed a ~500M-yuan round in early October. |
| LFM3 series (Liquid AI) | 10-27 | On-device multimodal (open weights) | Six ~2B open models at once: text, vision, audio, omni, plus two decision models, all branched from one 1.7B base pretrained on 22T tokens. The pitch is small: 4-bit quantized ~1GB memory, runs on phones and Raspberry Pi 5; vendor-claimed 1.8-2.8x faster edge decode; MacPaw measured 64% on 65 real macOS agent tasks (Qwen3.5-2B: 48%) with 58% fewer tokens. Note: mostly vendor-reported; small models still lose on big knowledge benchmarks - the vendor admits it. |
| Hy-SimulMT (Tencent Hunyuan) | 10-27 | Realtime simultaneous interpretation (open source) | End-to-end "speech in, speech out" interpretation: ~3s average latency, faster than the 4-5s human baseline; beats mainstream commercial SI APIs on the open RealSI long-audio benchmark; native zero-shot voice cloning preserves the speaker's timbre and emotional rhythm, Chinese-English. |
| Kimi CLI (Moonshot) | This week | Developer tool (open source) | A command-line coding assistant, technical preview: keeps native shell operation, Ctrl-K toggles an "agent mode" (context understanding, autocomplete, error explanation, natural-language script generation); native ACP support for IDE plugins and GUI clients. Moonshot's first open source on the developer-tool side; code on GitHub. |
| Cursor Agent SDK + DeepSeek V4.1 Flash | 10-27 | Coding-tool ecosystem | Cursor ships a code-first Agent SDK (Markdown prompts + TypeScript tools, self-host or Cursor Cloud) and adds DeepSeek V4.1 Flash to its catalog, vendor-claiming a new low of $0.60-0.98 per task on CursorBench. Closed tools are seriously adopting open cheap models - a line worth watching. |
Today's theme is "the bill war." OpenAI does flagship work at 1/10 the price, Anthropic opens its books to sprint for a $2T IPO, DeepSeek takes 150B yuan in industrial capital - money and capability are concentrating at the top, while open source keeps pushing the "good enough and cheap" ceiling lower. For users the advice is plain: stop staring at the #1 leaderboard spot; first figure out what each of your tasks actually costs - now is the best time to negotiate with a quote sheet in hand.
(Data from official announcements and public reports, 2026-10-28; 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