KI-Wochenrückblick Oktober 2026: Agenten-Harnesses, Scion vs. Git und neue Modell-Releases

Okay, Artificial Intelligence news updates every week Friday 2:00 PM Eastern time, New York time. Today is Friday, October 23rd. So, we have a lot of news as usual. The slogan for today is that the importance of harness, that harness even more important than the model and this picture explains the difference between Sian and Git. In Git, it's more like a mom-and-pop shop where each message is packed separately whereas Sian, it's more like a supermarket where you just grab stuff from the shelf. It's like self-help store and because of that, less friction and we'll talk about it. Okay. Next is the latest benchmarks. What's interesting that they are very, very close if you look at, for example, for coding, 1552, 1551, 1550, 1548, so it's very, very small difference between the top positions and Gemini Argon is only like 1. And never mind, okay. So, next is Artificial Analysis Intelligence Index. So we are talking about cost per doing some work. Uh GPT 61 SOL is a great choice because it's not expensive but this is a frontier model, I keep repeating about it, because look, on the right, we have some expensive Anthropic models which are 10 times more expensive. But also look, here GPT 6 Luna, and it is what 10 times less expensive than SOL and 100 times less expensive than Claude. Now we have Claude Haiku, which is fast, but really, not good. You can only use it for simple operations. But point is that you have some models which you can use in America for business. What what happened I'll talk about it. Our government, American government, they decided to block DeepSeek. Block, I mean that, projects, which are dealing with government, even if they're not military, if they're civil, they cannot use DeepSeek anymore. Uh so first, they blocked Alibaba, now they blocked DeepSeek. Anyway, uh Google's Harness 2 agent framework. Apache 2.0 code on GitHub, they have prompts, tools, memory, guardrails, skills, and other agent infrastructure without changing model weights. Task-level recursion uses Propose-Compose to compare models for task-level execution. Harnessing-level recursion preserves lessons to guide future tasks. Adaptable agent architecture rather than model retraining, with parallel or sequential refinement, OpenCode and Codex executors, and other related benchmarks. And Mistral from France releases Timestral models. Four sizes, 5M, 28M, 123M, and 209M parameters. Support multiple related time series, historical and known-future inputs, and probabilistic predictions that quantify uncertainty. They are fast, that's what, run on CPUs, Mistral, and Timestral and few-bench among pretrained zero-shot models, demonstrating competitive accuracy with fewer parameters. Gemini 4 Argon goes GA with a 1M-token output limit. Open to developers, enterprises and Gemini Pro and Ultra. One response can run to 1M tokens (up from 64K). Input and output share the window. API price $2 / $10 per 1M tokens until November 30, then $4 / $20. Google internally reports 35% more agentic code submissions and about 30% fewer rollbacks. Available in Cursor, Devin and GitHub Copilot. And Minimax launches M3.1-flash AI model. For coding, autonomous agents, computer ops, cybersecurity. It's 3 times faster, 70% lower agent-task costs than previous version. API pricing is $0.15 / $0.60 per 1M tokens in/out. Available via Minimax's API, M Plan, Minimax Code, Ollama. Model weights will follow later. DeepSeek V4.1-Flash-1017. DeepSeek V1.1 coding score increased from 74.2 to 77.0. It's 25% faster in inferencing, faster task completion, shorter outputs. Same price as before. DeepSeek Search and Fetch APIs - $0.75 per 1,000 calls. REA - Reverse Engineering Anything. Use AI and MCP with software-analysis tools to do reverse engineering and re-implementation of software. This started in April, 2024 and then it exploded in October - currently 371K stars. TypeSafe Jev 1.14 - 262K context, images + location tasks. Can accept image input, 262K text context. Can be used as a type - for object-location tasks. Developers can combine images with instructions and questions and code. Longer contexts reduce speed and accuracy. It works as a decision model - for 100K input tokens, with free output. 20 Jev alternatives, Cloudflare Clef / Def-flash - multimodal, open-weight decision models. OpenAic, Claude, Google, Gemini - host models for text, images, and audio. OpenAi Decisions API - hosted chips using 0.6 cents per 1M tokens. MeraGPT decider 1 - tested, text-only, jev-compatible model. LiteM - small, fast (like 262K context) and ability scoring. Standard Thinking Standard one - 38B and 10B text / image models. Together AI Tevi-4B-experimental - experimental multiple-choice-themed models. Fastio Gliner2.5-Decide - compact converter for decisions and extraction. Strands Labs Strands Decider - approximately 2B, self-hosted, with training recipe. Kev - jev-compatible models from Jared Palmer, 088-278. Decider Pack - self-hosted family, 0.8B-35B-43B. Bespoke Labs Bespoke Nimble-9B - lightweight adapter for text decisions. ModernBERT-Decide - a special model, ModernBERT based decider. Von - small encoder that scores all options in one pass. Open-Jev - open-based models on 2B, 9B, and 278 sizes. OpenDevice on DiffusionGemma - text/image model with jev-compatible API. Innerlev - 4B and 278 models distilled from LLM reasoning. Rizzo Flow - self-hosted model with native and jev-compatible APIs. Weekly videos every Friday. 7.65K subscribers, 375 videos. 1. Subscribe to this channel, https://www.youtube.com/@lev-selector. 2. Download slides from GitHub using links under the videos. 3. Please pause the video - and answer the pinned question in comments under the video. Scion vs Git. Anthropic reports a 200+ smaller CPU-and-memory infrastructure footprint than its previous Git mirroring infrastructure, and roughly 100+ faster checklists. Git often works like a shopkeeper assembling a custom shipment based on what each customer already owns. Scion works more like a supermarket with standardized, prepackaged items on nearby shelves: clients request items they need, they get needed, and shared caches hand those objects over in a ready-to-use format. This avoids overheads in a ready-to-use-to custom shipments. Antique: Anthropic's Code Review and Approval Tool. Antique addresses review and approval bottlenecks as AI-generated code overwhelms workflows designed for human oversight. Review tooling supports approval rules that distinguish between minor edits from sensitive logic changes, directing human attention toward consequential modifications. NVIDIA Nemotron 3.5 Super VL, a 120B open multimodal model for AI agents. Hybrid Mamba-Transformer, 1M-token context. Multimodal, handles text, images, documents, gui, video up to 2 min. NVFP4 build runs on one DGX Spark, H200 or GH200, or two H100s. NVIDIA claims tuned versions beat Claude Opus 5 on some tasks. Smaller Bolt 74B-A98 fits in 24GB; free endpoint on OpenRouter. OpenAI ships GPT-Live-1.1 with 13 new voices. ChatGPT Voice, Codex and the dots personal agent. New voices in French, German, Hindi, Japanese, Korean and Mexican Spanish. Better accents and so on. "Broken checking..." during web searches. Liquid AI d1-ultra tops the AA Decision Index. A multimodal decision model - text or images in. Answers are fast and cheap. Score 65 on the Artificial Analysis Decision Index, vs 58 for TypeSafe Jev 1.13. Developers can combine images with instructions, tool calling and safety classification equally. Priced at $0.042 per 1M input tokens on the Liquid API. Krea 3: Image editing in under 8 seconds. Image editing model built for production work. Finishes an edit in under 8 seconds. Costs $0.06 per image. Krea say it's 4x faster than Nano Banana Pro. Combines up to 8 reference images into one picture. Krea AI is in San Francisco, California. AI Personality Becomes Competitive. Artificial Analysis introduces AA Personality Bench: 900 questions assessing 30 traits, grouped into 8 indices, including warmth, imagination, transparency, and generic filler. Capable models can feel cold, sycophantic, or frustrating and this figure illustrates personality-driven engagement; OpenAI's agents face criticism for dryness. The broader achievement is making conventional character measurable, strengthening product differentiation. Customize Claude Code. 1Password for Agents. Lets agents log in without seeing passwords. The user approves each request. 1Password fills the credentials into the site without seeing the paswd. Gives agents account access while users keep control of their secrets. Claude's text output now carries an invisible watermark. Based on SynthTID-text, added to comply with the EU AI Act. Works by nudging choices among near-equivalent words. Adds no hidden characters or tokens. The mark carries no information about the user. Anthropic junior hiring went down. Since 2025 from 27% to 9%. In August its agents logged x6 times more hours than its staff. Employees report productivity gains, but more time goes to supervising agents. Meeting time rose about 20%; human attention is becoming the bottleneck. Thank You!