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.

Ant Digital shipped a "not-quite-an-LLM" model: LingDT-Decision-1.0. It doesn't write articles, do support chit-chat, or generate reports - it does exactly one thing: judge among the options you give it: is this a payment demand or a complaint, does this ticket need a human, which tool to call, what risk tier.
In plain terms: before, making an LLM decide "which department does this ticket belong to" meant it first composed a sentence in its head, then output text you had to parse with code - a detour that's slow, expensive, and often badly formatted. LingDT cuts the detour and returns the option or score directly. Same job, minus the "generate text" step, and both money and time drop.
Old decision models were single-choice - one highest-probability option. But real business needs are "say several things at once, in order." The vendor's example is telling: a user says "refund me, cancel auto-renew, and issue an invoice, but don't delete my account" - a single-choice model only catches "refund" and drops the other three.
These sound like engineering details, but they decide whether it fits a business pipeline. A single-choice model in a ticket system is like making people say only half of what they mean.
On Oct 22, OpenAI made its Decisions API generally available, aimed at the same pain point: 100-180ms responses, $0.10 / M input, cache hit $0.01, output free, up to 400 questions per request, three return modes (probability / option / rating).
But a crucial difference: OpenAI didn't train a new model for this - it reconfigured the existing GPT-6 Luna and sped up the pipeline. Its product lead said it's "the same model."
The two routes diverge clearly: OpenAI adapts a general model into a specialized API, while Ant trains a dedicated small model. The former is fast to adopt with an existing ecosystem; the latter is cheaper and more stable in theory, but you shoulder training and maintenance. Which wins depends on volume - at tens of thousands of calls a day the API is enough, at hundreds of millions a day that 90% saving is real money.
First, all scores are Ant's own - the 23 benchmarks are public but no third party has reproduced them, and the "composite index" is a self-defined 90% single-choice + 10% multi-choice weighting. Second, weights aren't open - DTMaaS only, and pricing after the Oct 30 free window is unannounced. Third, it can't replace a general LLM - no complex reasoning, multi-turn dialogue, or writing. Its role is "scooping the high-frequency small judgments out of the general model's bowl," so a system usually runs two models side by side, not one instead of the other. Before adopting, run it on your own ticket history - more useful than reading leaderboards.
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| WorkForge (Tencent Hunyuan + Fudan) | 10-23 | Agent training environments + fine-tuned models | A framework for mass-producing "practice workstations" for AI: 16,665 workspaces built from real repos, datasets, and web pages, covering 40 domains and 60 file types, ~42.3 files and ~76K tokens per trajectory. Acceptance is double-checked: a program verifies numbers, another model judges quality. Results: 35B-A3B's GDPval up from 45.5 to 73.6 (+28.1); the 27B build hits GDPval 82.4, beating Qwen3.5-397B (79.7), 15x its size, though still below GPT-5.6 Sol's 90.1. The telling detail: demonstration trajectories were generated with Zhipu GLM 5.3 plus Claude Code and OpenClaw, then used to fine-tune Qwen3.5 - Tencent's training-data factory hired someone else's model as its first teacher. |
| Ling-3.1-flash-1019 (Ant Bailing) | 10-22 | General model (score only) | Announced on X: 44 on the AA intelligence index, tied with Kimi K3, up 3 from Ling-3.1-flash's 41, just 1 below Qwen3.8 Max, DeepSeek V4.1 Flash 1017, and GLM-5.3(max). AA-Briefcase Elo ~1400\u21921524, AutomationBench-AA 62%\u219268%. But scores only - no weights, no API; AA lists no public API, proprietary license, undisclosed params, with weights/API "coming soon." Read it as a published report card, not a usable model. |
| Canva Craft | 10-22 | Image foundation model (in-house, closed) | Canva's first from-scratch image model, built for "directly usable design files." Two-stage: a planner model writes the composition (elements, placement, size, text), a renderer draws it - up to 2K, 30+ aspect ratios, in-image text in 55 languages. Trained not on photos but on millions of hand-made design templates (posters, menus, PPTs, infographics). DesignArena image #4 (1329) - the top three are all OpenAI models, 38 points above Google Nano Banana 2.1. Open to all (Dream Lab), API as craft_v1_2k (preview). Vendor-claimed ~30x lower cost - unaudited. |
| Liquid RIPL + LFM2.5 | 10-22 | Realtime inference engine | Liquid AI's realtime inference pipeline, aiming to put language models directly into the request hot path. On one B200, the 1.2B LFM2.5 hits p95 34.6ms (1000 req/s), vs 0.8B Qwen3.5 on vLLM at 498ms - a bigger model, 14.4x faster; under load the gap widens to 28.4x (155.76ms vs 4430.88ms). Shopify already uses it in production search augmentation at under 20ms. The point: displacing gradient-boosted trees from risk, ranking, and payment-auth slots with semantically sharper language models. |
| ViGAR (Zhiyuan x PKU) | Oct | World-action model (full-size humanoid) | Claims the first complete world-action model for full-size bipedal humanoid robots, verified on the Zhiyuan AgiBot A2 for multi-step tasks (clearing table clutter into a bin, plating buns from a double steamer). RoboTwin 2.0 Clean2Random: Clean 82.00%, Random 67.02%, avg 74.51%, +12.81 points over the strongest baseline; topped at launch, now #3 overall as the board updates. Same batch: Zhiyuan open-sourced the AGIBOT WORLD 2026 Phase-4 "embodied reasoning" dataset. |
| mammoth-1 / elephant-1 (Base Labs) | 10-21 | Decision models (open weights) | Same lane as LingDT: Qwen-based open-weight decision models in 27B and 9B, returning calibrated probabilities rather than generated answers. Vendor claims mammoth-1 leads the Jev decision index. Read with Ant's model: "decision model" is now a category, not one vendor. |
| OlmoEarth v1.3 (Ai2) | 10-22 | Remote-sensing embeddings (fully open) | Ai2's earth-observation embedding update, weights + pre-training data + code together, under the OlmoEarth Artifact License. 10m resolution, 64-dim and 128-dim builds, the 128-dim tops 20 of 27 evals. Ships global precomputed embeddings and a browser explorer - usable without training. |
| Chariot-1-ASR | 10-22 | Realtime speech recognition | An Indian startup's realtime transcription model for 10 Indian languages, vendor-claimed lower error than Google and Sarvam. License and availability undisclosed; a signal that "Indian-language realtime ASR" now has dedicated players. |
Every headline today points to the same thing: large models are being taken apart into components. Ant split "judgment" out into a 4B/27B part, OpenAI repackaged a general model as a decision-only API, Liquid split "speed" out with a dedicated engine, Canva split "layout and text" out and trained on design files, and Tencent split "the training ground" out with 16,700 workspaces - so a 27B model beat a 397B. Six months ago everyone was comparing who had the stronger model; now it's "which job gets sent to whom." For model users the conclusion is concrete: stop running everything through one big model; sort your traffic into "hard" and "high-volume," and the high-volume part can save more this year than any price cut.
(Data from official announcements and public reports, 2026-10-23; 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