AI news: January 2026

January 2026 was a month of agents and open weights. The headline is Cowork, Anthropic’s desktop agent that moves Claude out of the chat window and into the user’s local folders. Around it, OpenAI moved into health with ChatGPT Health, and open models (FLUX.2 klein, NousCoder-14B, Kimi K2.5) kept pushing prices and latency down.

Anthropic ships Cowork, a Claude Desktop agent for local files

Anthropic announced Cowork on January 12 as a research preview: a new tab in the Claude desktop app, next to Chat and Code, that gives the model permission-based access to local folders so it can read, edit and create documents without manual uploads. Given a task, the agent drafts a plan, runs steps in parallel, reviews its own output and asks when it gets stuck; several tasks can be queued at once. At launch it was limited to Claude Max subscribers ($100-200 per month) and macOS only.

For a team already on Claude, this extends the Claude Code pattern, MCP servers included, to everyday office work: reports, spreadsheets and documents that live on disk, with no manual integrations in between. It remains a research preview with limited access.

Source: VentureBeat · also on Anthropic and Simon Willison

OpenAI launches ChatGPT Health

OpenAI introduced ChatGPT Health on January 7: a dedicated space that connects medical records and wellness apps (through connectivity partner b.well, plus Apple Health, MyFitnessPal, Weight Watchers and Function) so answers reflect the user’s clinical context. Health conversations get their own encryption and isolation and, per OpenAI, are not used to train its base models. The company says 230 million users ask about health every week. Initial access is limited to early users and currently excludes the European Economic Area, Switzerland and the UK.

For digital-health product teams it shows where the sector’s biggest vendor sets the bar: real clinical-data integration with a separate privacy treatment.

Source: OpenAI · also on TechCrunch and CNBC

Grok Imagine API: xAI’s video generation for developers

xAI opened the Grok Imagine API on January 28, bundling video generation and editing. The model produces clips of 1 to 15 seconds at 24 fps in 480p or 720p, from a text prompt, a still image or up to 7 reference images, and it can edit or extend a generated scene to restyle it or add and remove objects. Per xAI’s documentation, generation costs on the order of $0.08 per second of video. The Artificial Analysis evaluator ranked it top in text-to-video, well below Sora’s price.

For video production teams, it is another aggressively priced high-end generation API, useful for mockups and short pieces; 15 seconds is the ceiling per clip.

Source: xAI · also on Latent.Space and OpenRouter

FLUX.2 klein: sub-second image generation with open weights

Black Forest Labs released FLUX.2 klein on January 15: an image-model family in 9B and 4B sizes (plus undistilled Base versions and FP8/NVFP4 quantizations), with inference under 0.5 seconds according to the maker. The 4B runs in about 13 GB of VRAM, which means it fits on consumer GPUs, and ships under Apache 2.0; the 9B carries a non-commercial license. An independent test by 302.AI confirms the speed but flags weaknesses in realistic textures and prompts requiring precise spatial logic.

For anyone generating images in production, an open model at this speed on your own hardware changes the cost math against closed APIs, though the 9B license and the independently reported quality limits remain caveats.

Source: Black Forest Labs · also on Hugging Face and 302.AI Benchmark Lab

NVIDIA Cosmos Reason 2 brings visual reasoning to physical AI

NVIDIA published Cosmos Reason 2 on January 5 in 2B and 8B sizes, expanding the context window from 16K to 256K tokens and adding OCR, 2D/3D grounding and trajectory data support. NVIDIA says it leads Physical AI Bench among open models. Uber uses it to caption autonomous-driving video, and Salesforce, Hitachi and VAST Data appear as early adopters in robotics and video analytics. The models are on Hugging Face; deployment on AWS, Google Cloud and Azure was announced but is not yet available.

Relevant for teams working with industrial video, robotics or visual inspection: an open VLM that reasons about real-world physics and can be fine-tuned on your own data. The cited benchmarks are the vendor’s own.

Source: NVIDIA on Hugging Face · also on VentureBeat and NVIDIA docs

Veo 3.1 Ingredients to Video: consistent characters across shots

Google DeepMind updated Veo 3.1 on January 13 with “Ingredients to Video”: up to 4 reference images to keep the same character, objects and color palette across takes, native 9:16 vertical video, and 1080p/4K upscaling (the latter only in Flow, the Gemini API and Vertex AI). Videos carry a SynthID watermark that the Gemini app can verify. Third-party sources put API pricing at $0.15 per second for the Fast variant and $0.40 for standard, with rollout centered on the US for now.

Character consistency was the usual blocker for using generated video in anything longer than one shot; with this update a content team can consider short branded sequences without a shoot.

Source: Google · also on YouTube support

Luma’s Ray3.14: native 1080p, four times faster

Luma Labs shipped Ray3.14 on January 26: per the company, 4x faster and 3x cheaper than Ray3 at 720p, with native 1080p in Dream Machine and up to 18-second clips in the Modify flow. Luma itself lists the limits: no character references and no HDR/EXR in Modify Video. Third-party reviews highlight photorealistic human motion and camera coherence, and confirm those same gaps.

For AI video production, price and wait time dropping together means more iterations per budget; the missing character reference keeps it a step behind Veo for series built around a fixed protagonist.

Source: Luma Labs · also on Luma pricing

NousCoder-14B: open-source coding model trained in four days

Nous Research released NousCoder-14B, a coding model built on Qwen3-14B and tuned with reinforcement learning over 24,000 verifiable programming problems: generated code gets executed and scored. Training took about four days on 48 GPUs. It reaches 67.87% on LiveCodeBench v6, 7 points above its base model, per Nous Research’s figures. The company also published the full RL environment, the benchmark set and the training harness on its Atropos framework.

More than the model itself, the usable part is the published recipe: a team with GPUs can reproduce the process on its own coding problems.

Source: VentureBeat · also on Nous Research

Kimi K2.5 adds vision and agent swarms

Moonshot AI launched Kimi K2.5 on January 27, trained on some 15 trillion mixed vision-and-text tokens. It comes in four modes (Instant, Thinking, Agent and Agent Swarm, the last in beta): the swarm coordinates up to 100 subagents and up to 1,500 tool calls, cutting execution time by up to 4.5x per Moonshot. The company cites 76.8% on SWE-Bench Verified. Pricing, $0.60 per million input tokens and $2.50 per million output, sits far below Western frontier models. An independent review reports very low safety scores without a system prompt and frequent hallucinations in real use.

It is the one to watch for large agentic workloads where per-token cost dominates; the independent review’s safety and reliability caveats argue for running your own evals before anything serious.

Source: Moonshot AI · also on GitHub and Awesome Agents

Mistral Vibe 2.0: a terminal coding agent on Devstral 2

Mistral introduced Vibe 2.0 on January 27, its terminal-native coding agent powered by Devstral 2 (123B, 72.2% on SWE-bench Verified per third-party benchmarks). It adds custom subagents, clarifying questions before executing, and slash-command skills. Devstral 2 costs $0.40 per million input tokens and $2.00 per million output. In Mistral’s own human evaluation it beats DeepSeek V3.2, though the company concedes Claude Sonnet 4.5 remains clearly preferred. Reviews find it solid on scoped tasks and prone to losing the thread in complex multi-file work.

For European teams with data-sovereignty requirements it is the most complete local alternative to the US coding agents, with the tempered expectations Mistral itself admits.

Source: Mistral AI · also on Awesome Agents