Gemini 4, GPT 6.1, Plasma conjecture, Napoleon code solved, OpenAI Dots, Claude Sonnet 5.5: AI NEWS
AI Search showed 26 AI tools in this video on . 16 are verified available in the EU. Each row below jumps to the exact second. Part of week 2026-40.
Tools shown in this video, with timestamps
Takes an image, multiple images, panorama or video plus a camera path and generates a video walking through the scene. Code released, based on Wan 2.1 1.3B, under 6 GB.
Upscales low-res video to up to 4K in a single step and is model agnostic. Output is sharper but changes details, so it is not faithful. Code released, borrowed from LTX Refiner.
A 16.9 MB speech-to-text model that runs on CPU with no dependencies. It handles 30 seconds of audio, auto-detects language, gives word timestamps, and has lower error rate and faster decoding than Whisper Base.
Live demoAn open-source speech-to-text model under 900 MB that transcribes an hour of audio in about 20 seconds. It has the second-lowest average error rate versus much larger models.
Live demoAnthropic's new smaller model, close to Opus 5.5 on agentic coding and knowledge work, and now the default on the free plan. It ranks second on the Artificial Analysis index but uses the most output tokens per task, making it the most expensive to run.
A persistent AI agent with its own always-on cloud computer and browser, connecting to apps via plugins and working across ChatGPT, Slack and Teams. Powered by GPT6 Astra, rolling out to Pro and business users outside the EU.
A self-hostable open-source alternative to OpenAI's Dots, released hours later. It is a framework to which you link your own or a local LLM.
An upgrade of GPT6 Sol, claimed to be near Astra intelligence for a fifth of the price. It is one point behind Astra on the Artificial Analysis index and is the most cost-efficient frontier model. It is available to paid plans in GPT work and Codex. Creator also notes the Pro plan usage cut and a new $500 plan.
An agent built into ComfyUI that plans steps, creates and connects nodes, runs workflows and can fix errors from a prompt. Currently available in Comfy Cloud, with Comfy Desktop support coming in a few weeks.
A unified model that understands and generates images and videos directly in pixel space, with no latent space encoder or decoder. Quality is below frontier models, as this is a research preview. Code and model are released under mostly Apache 2.
An image editing model billed as the most precise edit model, making pixel-level consistent edits through natural language and bounding box regions. It stays consistent across many edits, unlike GPT Image or Nano Banana. Open-sourcing is planned.
A closed-source image model with bounding-box composition control, text-to-image, up to 10 reference images and natural language editing. Weights are promised to be open later, but the creator notes Flux 3 video was promised the same and still isn't open.
Google's new frontier model that on average beats GPT6 Astra and top Claude models on benchmarks and ranks first on several. It can output 1 million tokens and has the lowest hallucination rate (15%) on the Omniscience benchmark. Access is currently limited to trusted cyber defenders, with paid API and Ultra rollout next.
A distillation method that lets video models like Minimax generate in four steps instead of 20-30, using discriminator heads to compare teacher and student outputs. Code released with LoRAs of 1.4 GB each.
Another four-step distillation method that projects away the critic's errors from the training signal. It beats other distillation methods, with a full 66 GB modified model and a 1.4 GB LoRA released.
Two papers used GPT6 Astra to find counterexamples to a plasma physics question open since 1967, showing stable smooth 3D plasma configurations can exist without symmetry.
A system giving humanoid robots touch sensing via pressure insoles, trained on the Unitree robot. It reduces peak touchdown force by almost 50% and adds up to 24% more contact area.
A Unitree G1 humanoid trained to autonomously play badminton, including forehands, backhands and jumping returns, rallying with a human. Trained on synthetic data expanded from limited real hitting motions.
ElevenLabs' latest text-to-speech model, described as the most emotive, with metatags to control direction, sound effects and tone. It is on the free plan and tops a TTS leaderboard.
Live demoCarter Church used GPT6 Astra to decipher a 217-year-old coded Napoleon letter. The model recognized about 1,300 cipher units and built a solver that reconstructed the key in roughly 6 hours. The letter was a military briefing to General Marmont.
Generates 3D human animation of a person interacting with an object from a text prompt. Body parts vote on object motion and a refiner cleans up contact points. GitHub is out under MIT, with models to come.
Takes an image or 3D mesh plus point prompts and decomposes the object into non-overlapping parts without gaps. It beats similar segmentation tools on benchmarks. Models and training code are to be released.
An 8B model based on Qwen3 that turns a prompt plus papers into a cited research report in one pass. It beats base Qwen3 and is close to Dr. Tulu, with model and data fully open, 32 GB in size.
An open-sourced training system for mixture-of-experts models that keeps each expert on its own GPU and routes work to it. It makes training about 2.7 times faster than the previous system.
An agentic MoE model for CLI tools like Claude Code and Codex, with 320B total and 15B active parameters. It works through long software tasks and performs well for its size. Code is released.
A 27B Qwen-based model trained to iteratively propose, test and refine its own solutions on verifiable ML and algorithmic tasks. It matches much larger models, including Kimi K3 and GPT 5.6 Sol on some benchmarks, and is a 55 GB download.
