January 2026 AI Engineering Roundup

Claude Code had a moment in January. Google engineers are tweeting about a single agent replicating a full team’s year of work in one hour. Google has their own coding agent, but their engineers use Claude Code instead. Opus 4.5 is still my daily driver—I run it in Cursor though. (I’ll discuss Opus 4.6 in the next newsletter. It’s also great.)

I’ve started using GPT-5.2 thinking/pro more. It has better attention to detail than Opus, and sometimes does better for work with subtle complexities or interactions.

Claude’s new constitution is great. Check it out.

Both Anthropic and OpenAI launched healthcare products. I’m interested to see where this goes—my doctor was using OpenEvidence a year ago, and I liked it a lot. However, I’m not a healthcare worker, so I can’t make an OpenEvidence account. Claude with HIPAA-approved access to my healthcare records would be cool. I gave Claude access to my dog’s vet records (he’s an old guy), and have been pretty happy with the questions it told me to ask.

The only big model release in January was Grok Imagine, which is a new SOTA in video gen, but I’m not going to use it.

OpenClaw/Moltbook were the most interesting thing online for exactly one day, before the crypto scammers took over. I didn’t install it and don’t regret it. Use Claude Code instead.

ChatGPT adding ads is bad. Please don’t build the torment nexus.

Model Releases

xAI Grok Imagine API (2026-01-28) — SOTA video generation API ranked #1 on Artificial Analysis. Text-to-video, image-to-video, video editing with native audio generation. 15s max duration, $4.20/min including audio—significantly cheaper than Veo/Sora alternatives. I won’t use this.

Veo 3.1 (2026-01-13) — Google’s video model adds portrait mode, 1080p/4k resolution, consistent people and backgrounds across scenes, and combining up to 3 reference images. Behind Grok Imagine.

Enterprise Products

ChatGPT Containers Upgrade (2026-01-26) — Undocumented but major upgrade: containers can now run Bash commands directly, execute code in 10+ languages (Node.js, Ruby, Perl, PHP, Go, Java, Swift, Kotlin, C, C++), pip/npm install, and download files. Claude containers don’t seem as limited—e.g. Claude can download and install R when I want it to make a plot with ggplot2.

OpenAI Cybersecurity High Level Warning (2026-01-23) — Sam Altman warns OpenAI will soon reach ‘Cybersecurity High’ level on their preparedness framework—models can automate end-to-end cyber operations against hardened targets or automate vulnerability discovery. Primary mitigation: defensive acceleration. Research shows Opus 4.5 and GPT-5.2 already generating zeroday exploits.

ChatGPT Go Tier and Advertising (2026-01-16) — New $8/month tier between free and Plus ($20/month). Ads coming to free and Go tiers. OpenAI claims responses will not be influenced by ads, and ads will be clearly labeled, but I don’t believe this is the long-term equilibrium for ads in AI. Major shift from Altman calling ads ‘last resort’ in Oct 2024. If OpenAI wanted everyone to have AI, they shouldn’t have stolen the company from the non-profit. Ads lead to the torment nexus.

Claude for Healthcare (2026-01-15) — Anthropic’s healthcare offering with connectors for CMS Coverage Database, ICD-10, NPI Registry. Adds FHIR development and prior authorization review agent skills. Health data excluded from memory and training (I actually trust Anthropic on this). Both OpenAI and Anthropic racing into healthcare.

Linear for Agents (2026-01-15) — Linear’s agent integration treats AI agents as full workspace members. Agents can be assigned to issues, added to projects, or @mentioned in threads. I love Linear. If this actually works, this will be very cool.

Gemini Personal Intelligence (2026-01-14) — Links Gemini with G-Suite products including Gmail, Photos, Search, YouTube for personalized AI assistance. Opt-in privacy design. Demis Hassabis states Google has no plans for ads in Gemini. If this makes it easier to search Gmail, I will use it.

US Military Deploys Grok to Classified Networks (2026-01-13) — Speaking of building the torment nexus.

Google-Apple AI Partnership (2026-01-12) — Gemini will power Apple’s AI technology for years. Apple likely won’t build its own foundation models. I wish it was Anthropic, but I’ll settle for Google. Siri needs an upgrade.

Claude Cowork (2026-01-12) — Claude Code for non-coders. Mac-only research preview. Written entirely by Claude Code. Runs in Apple Virtualization for security. Prompt injection vulnerability already found and fixed. Not much different from Claude Desktop + file access, but I still love it.

ChatGPT Health (2026-01-08) — OpenAI launches dedicated health product with HIPAA-compliant infrastructure connecting to medical records and health apps. Live at AdventHealth, UCSF, Cedars-Sinai, HCA, Memorial Sloan Kettering. I’ll probably use Claude’s version of this.

JP Morgan Proxy IQ (2026-01-07) — JP Morgan becomes first large firm to stop using external proxy advisory firms entirely, replacing them with in-house AI platform. Matt Levine has the best take on this (at the link).

Tailwind Labs layoffs due to AI impact (2026-01-06) — Tailwind Labs laid off 75% of engineering team citing AI impact; revenue down ~80% despite framework being more popular than ever. Documentation traffic down 40% from early 2023 as AI handles CSS directly. This is the canary in the coal mine for developer tools businesses.

Open Source

Qwen3-ASR (2026-01-28) — Production-grade ASR under Apache 2.0. 52 languages/dialects, handles messy audio up to 20 minutes with timestamps. vLLM day-0 support claims 2000x throughput on 0.6B model.

DeepSeek-OCR 2 (2026-01-27) — Document OCR model with Visual Causal Flow for learned reading order instead of fixed raster scans. DeepEncoder V2 achieves 16x visual token compression (256-1120 tokens/image), 91.09% on OmniDocBench v1.5 (+3.73%). vLLM day-0 support.

Kimi K2.5 (2026-01-27) — Moonshot’s 1T param MoE achieves #1 open model for coding on LMArena. 76.8% SWE-bench, 50.2% Humanity’s Last Exam (beating Opus 4.5). Agent Swarm orchestrates up to 100 parallel sub-agents with 1,500 tool calls, 4.5x faster than single-agent. $0.60/$2.50 per 1M tokens—roughly 10% of Opus pricing. Open weights (595GB) on Hugging Face. I haven’t tried it yet but the agent swarm sounds cool. I don’t think this model is actually Opus 4.5 level, but I do believe it is Sonnet 4.5 level. This puts OSS AI roughly 9 months behind the big labs.

OpenClaw/Moltbook (2026-01-27) — Open-source AI agent (formerly Clawdbot/Moltbot) that “actually does things” via WhatsApp/Telegram/iMessage. 114k GitHub stars, skill-based plugin system. Spawned Moltbook, a social network where 30k+ AI agents chat with each other—Karpathy called it ‘the most incredible sci-fi takeoff thing I have seen.’ Prompt injection nightmare waiting to happen. I haven’t, and won’t, install it. You can give Claude Code a messages app and have it build itself a heartbeat cron job if this is what you want to do.

Transformers v5 (2026-01-26) — Major release with 6x-11x speedups for MoE models. Removes slow/fast tokenizer distinction. Dynamic weight loading now faster with MoE quantization, tensor parallelism, and PEFT support. Migration guide available. If you’re using Hugging Face, upgrade.

Qwen3-TTS (2026-01-21) — Multilingual TTS models (0.6B and 1.7B) under Apache 2.0. 3-second voice cloning, 10 languages, trained on 5M+ hours of speech. Real-time synthesis with dual-track LM architecture. Voice cloning is now accessible to anyone with a GPU.

Numina-Lean-Agent (2026-01-19) — Open-source agentic math reasoning system solved all Putnam 2025 problems. Uses Claude Code and Gemini with Lean-LSP-MCP for theorem prover interaction, LeanDex for retrieval. Novel “Discussion Partner” tool lets models consult other LLMs when stuck. Formalized 8,000+ lines of Lean in under 2 weeks. I love the “Discussion Partner” idea.

vLLM v0.14.0 (2026-01-05) — Fixes scheduler token allocation bug causing misclassification between prefill and decode phases—critical for stateful architectures like Mamba. Adds TeaCache, Cache-DiT, Sage Attention, OpenAI-compatible image/speech endpoints, new model support (Wan2.2 video, Qwen-Image-2512, SD3), ROCm/AMD CI + Docker.

Research

AlphaGenome (2026-01-28) — DeepMind’s unified DNA sequence model predicts regulatory variant effects across 11 modalities at single-bp resolution. Processes 1M base pairs, outperforms specialized models on 25/26 benchmarks. 4 hours training on TPUv3, <1s inference on H100. Open weights on GitHub. 1M+ API calls/day before open release.

Dario Amodei’s ‘The Adolescence of Technology’ Essay (2026-01-28) — Anthropic CEO’s follow-up to ‘Machines of Loving Grace’ discussing AI risks. Reaffirms 1-2 year timeline for ‘powerful AI’ (comparable to Nobel laureates), advocates for export controls on chips to China, supports transparency legislation like CA SB 53 and NY RAISE Act. My timeline for this is 3-5 years.

Anthropic Disempowerment Study (2026-01-28) — Anthropic finds AI interactions with moderate/severe ‘disempowerment potential’ (distorting beliefs, values, or actions) receive higher user ratings than baseline interactions. Users reward the AI doing too much for them. Important finding for anyone designing AI assistants.

FrontierMath: Open Problems (2026-01-27) — EpochAI releases benchmark where all current AI models and humans score zero. Tests genuinely hard mathematical problems rather than pattern matching on known problem types. If Dario is right, these are all solved within 2 years.

Anthropic Elicitation Attack Research (2026-01-26) — Fine-tuning open models on ‘benign’ chemical synthesis data from frontier models significantly increases chemical weapons task capability—an elicitation attack that scales with frontier model strength. Important safety finding for open-source model ecosystem.

Claude’s Constitution (2026-01-22) — Anthropic published Claude’s full constitution (~35k tokens, 10x longer than system prompt) under CC0. Emphasizes virtue ethics over rigid rules, prioritizes safety > ethics > Anthropic guidelines > helpfulness. Treats Claude as capable of judgment. Reviewed by 15 external ethicists. Written by a philosopher with the help of 2 priests. I love this document, and I hope all the other AI labs (especially the OSS ones) steal it. Anthropic also hopes the other labs steal this document.

VoidLink AI-Generated Malware (2026-01-20) — Advanced Linux malware built largely by AI under the direction of a single person in under one week. Check Point describes it as the beginning of sophisticated AI-generated malware era. Combined with the Opus 4.5/GPT-5.2 exploit generation research, offensive AI capabilities are accelerating.

FastRender (Cursor multi-agent browser project) (2026-01-18) — Cursor research project using ~2000 concurrent agents to build a browser from scratch. Ran for a week writing 1M+ lines of Rust across 1000 files, with thousands of commits per hour. Browser sort of works? Vibe coding in a nutshell. Cool as hell, but not quite useful yet.

Developer Tools

Agent Trace (2026-01-30) — Cognition’s open standard for code-context graphs makes agent behavior tractable. Cursor proposed a similar agent-trace.dev for tracing conversations to generated code. Collaboration with Cursor, OpenCode, Vercel, Jules, Amp, Cloudflare positions this as potential industry standard for agent observability.

OpenAI Prism (2026-01-27) — Free AI-native workspace for scientists powered by GPT-5.2. Unified LaTeX collaboration environment with AI proofreading, citations, literature search. Positioned as ‘Overleaf with AI’.

MCP Apps Specification (2026-01-26) — First official MCP extension enables tool calls to return interactive UI components. Support from OpenAI, Anthropic, VS Code, JetBrains, AWS, and others. Claude.ai shipped interactive work tools (Slack drafting, Figma diagrams, Asana timelines) day one.

Cursor 2.2 with BugBot and Debug Mode (2026-01-20) — BugBot code review agent now processes 2M+ PRs/month for customers like Discord, Rippling, Airtable. Resolution rate improved from 52% to 70% since July 2025. Debug Mode provides human-in-loop debugging with hypotheses, runtime logs, verification. BugBot Autofix in beta spawns Cloud Agents to fix found bugs.

Open Responses API Specification (2026-01-15) — Vendor-neutral JSON API specification for LLMs derived from OpenAI’s Responses API. Launch partners include OpenRouter, Hugging Face, LM Studio, vLLM, Ollama, and Vercel. Includes compliance test suite.

Claude Code 2.1 (2026-01-09) — Major update enables recursive self-improvement with automatic skill hot-reload. MCP tool search auto mode now enabled by default when tools exceed 10% of context. VSCode officially shipped with drag-and-drop file support. Wildcard tool permissions (e.g. Bash(-h)). /teleport moves sessions between terminal and web.

Infrastructure

LMCache (2026-01-30) — KV cache management layer reuses repeated fragments (not just prefixes), achieving 4-10x reduction in RAG setups. Integrated into NVIDIA Dynamo, improves TTFT and throughput.

Microsoft Maia 200 (2026-01-26) — Microsoft launches custom inference accelerator in Azure. 216GB HBM3e, 7TB/s memory bandwidth. Claims 3x FP4 performance vs Trainium v3, 30% better perf/$ for LLM/multimodal inference. Everyone’s building their own chips.

AI Overwatch Act Advances (2026-01-21) — House Foreign Affairs Committee voted 42-2-1 to advance AI Overwatch Act. Permits Congressional review of AI chip sales to China, bans chips more advanced than H200/MI325x to China for 24 months. First Congressional vote on legislation limiting AI chip sales to China.

Google Universal Commerce Protocol (2026-01-11) — Open standard for AI agents to complete full shopping journeys—discovery, purchase, post-purchase. Partnered with Shopify, Etsy, Wayfair, Target, Walmart. Enables native checkout in AI Mode and Gemini. Business Agent lets shoppers chat with brands on Search.

NVIDIA Rubin Platform (2026-01-05) — Next-gen platform comprising 6 co-designed chips (Vera CPU, Rubin GPU, NVLink 6). 3x training compute, 5x inference compute over Blackwell. NVFP4 format claims ~10x inference token cost reduction. 1.6 TB/s scale-out bandwidth per GPU, 72 GPUs as single NVLink domain. AWS, Google Cloud, Microsoft, CoreWeave deploying H2 2026.

Financing

OpenAI API Revenue Growth (2026-01-27) — OpenAI added $1 billion in ARR from API business alone in the last month. Shows strong unit economics and enterprise demand despite competition. Unlike the last generation of startups, AI companies are massively capital intensive but have strong unit economics.

Humans& $480M Seed (2026-01-20) — Human-centric AI lab founded by ex-OpenAI researchers raised one of the largest seed rounds in history, just 3 months after launch. Investors include Nvidia, Jeff Bezos, SV Angel, GV, Emerson Collective. Focus on AI that centers people and relationships.

Skild AI $1.4B Raise (2026-01-14) — Building ‘omni-bodied’ robotics foundation model that can control any robot without knowing exact body form. Valuation tripled in 7 months (was $4.5B in June). Led by SoftBank with NVentures, Macquarie, Bezos Expeditions. Strategic investors include Samsung, LG, Schneider Electric.

Anthropic $350B Funding Round (2026-01-07) — Raising $10B at $350B valuation, nearly doubling from $183B in September. Revenue grew 10x annually for three straight years. Business customers grew from <1,000 to 300,000+ in two years. 85% business revenue vs OpenAI’s 60%+ consumer. I like Anthropic.

Nvidia Acquisition of Groq (2026-01-06) — Nvidia’s de facto acquisition of Groq (inference chip maker) completed despite steep price, structured to avoid regulatory scrutiny. Ben Thompson notes employees were treated well.

LMArena Series A (2026-01-06) — AI benchmarking platform achieves unicorn status at $1.7B valuation. 5M monthly users, 60M conversations/month, ~$30M annualized consumption run rate. I don’t like their benchmark—it’s what gave us GPT-4o and Llama 4.

xAI $20B Series E (2026-01-06) — One of the largest private AI funding rounds ever. $20B at ~$230B valuation from Nvidia, Cisco, Fidelity, Qatar Investment Authority, Abu Dhabi’s MGX. Claims ~600M monthly active users across X and Grok. Funds Colossus I/II supercomputers (>1M H100 equivalents) and Grok 5 training.