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 Nov 5 (US Eastern), OpenAI formally released GPT-6 Cyber, a cybersecurity-specialized model open only to vetted Daybreak Red customers, and raised its cybersecurity risk rating under the "Preparedness Framework" from High to Critical - the highest tier. Reuters, ITHome, InfoQ, and the official release all agree: this is the hardest story today.
It's not "a general model with a security filter" - it's a product-grade model trained specifically for vulnerability discovery, exploitability verification, attack-path analysis, and patch generation. Key figures (all vendor claims, no third-party replication at publish):
Take: this is the first product-grade "offensive" model with a real-world track record, and the vendor itself raised the risk rating - a signal that AI security capability has crossed from "tool" to "weapon that needs guarding."
| Model / Product | Date | Type | Highlights |
|---|---|---|---|
| Kimi Hosted Agents (Moonshot) | 11-5 | Managed agent service | Turns the agent Harness into a cloud service: a 7x24 managed environment without self-hosting, long-horizon tasks and multi-agent collaboration; 140+ skills, 30+ plugins, per-agent sandbox + network control + session kill + audit export; Kimi Inference Hook blocks sensitive data before inference per enterprise rules. First 8 finance-agent templates; SF Express and CSC Financial onboard; enterprise-verified beta only. The pitch moves from "selling tokens" to "selling a runtime." |
| MNN-Pocket-0.3B (Taobao MNN team, open source) | 11-6 | On-device small model | ~300M params, 28-layer dense, vocab trimmed 151K to 64K, 4.1T tokens four-stage pretraining, 32K context. INT4 package just 173MB, peak memory +228MB (49% and 47% under Qwen3-0.6B). Real devices: iPhone 18 Pro ~350 tokens/s fully offline; iPad Pro M5 prefill 8721 / decode 466; Snapdragon 8 Gen 5 251. IFEval 73.20, Chinese IFEval-zh 74.59, 11-metric mean 50.00. Vendor-measured; aimed at "native in-app," not a general assistant. |
| Ripple-65B Preview (Ricursive, Apache 2.0) | 11-4 | Open weights (dense) | 65B fully dense, 80 layers, 12T tokens from scratch, claimed 1M context (hosts list 128K). Vendor: Terminal-Bench 2.1 80.52, above Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4) - hence "US open-source coding #1," but only 0.4 points above second. Its own card also lists SWE-bench Verified 76.5 (below Beam's 81.6) and DeepSWE 1.1 44.3 (far below Muse's 60.0). Total cost $32M, 1.4M RL rollouts (~1/70 of Beam), team grew 1 to 5; API $0.20/$0.60, NVFP4 runs on one DGX Spark. Preview, light post-training, all self-reported. |
| SWE-2 Flash (Cognition) | 11-4 | Closed coding model | Post-trained on Thinking Machines' Inkling-Small (276B-A12B), first to insert a mid-training stage before large-scale RL: Terminal-Bench 4.0 from 0.9% to 20.0%, DeepSWE 1.1 from 17.4 to 60.1, FrontierCode 2.0 24.0, but TB2.1 89.2 still below SWE-2's 92.8. Vendor: 3x faster, 1/5 the price, free for Devin users through Nov 30. Weights closed - targeted safety training lifted ExploitBench 11.3% to 34.8%, and the company cites cyber-risk for a partner-only whitelist. |
| PLaMo 3 Translate (Preferred Networks, open weights) | 11-5 | Translation-specialist | Four tiers 0.8B / 2B / 8B / 31B, 53 languages, fully offline on phone, laptop, or intranet (the 0.8B demo translates in airplane mode). ~37.3B tokens, SFT + DPO + small-model distillation; BF16/FP8/W4A16/NVFP4 and GGUF/MLX formats. Vendor: the smallest 0.8B already beats Google TranslateGemma and Tencent Hy-MT2 (30x its size). PLaMo Community License: free commercial under 1B yen annual revenue. Vendor self-assessment. |
| structRFM (USTC, open source) | 11-5 | RNA foundation model (AI for Science) | Bakes secondary-structure knowledge into large-scale sequence pretraining: ~21M sequence-structure pairs, structure-guided masking learns both sequence context and structure dependence. Sequence input alone yields sequence-, nucleotide-, and base-pair-level features. Landed in Nature on Nov 5 (DOI 10.1038/s41586-026-11210-2), USTC's 16th CNS paper in 2026; paired data, weights, and code all open. |
| Qwen3-30B-A3B-MoE-SAE (ByteDance Seed, open source) | 11-5 | Interpretability tool | TopK sparse autoencoders (k=100, 2^20 latents) trained on Qwen3-30B-A3B's MoE block outputs, layers 4/30/46, ~17.2GB per SAE, with a paper on "MoE experts exhibit human-interpretable fine-grained specialization" - a magnifying glass on what each expert manages. Apache 2.0 on HF. |
AI's offensive security capability is now real enough to require its own risk tier.
(Data from official announcements and public reports, 2026-11-06; 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