Mistral Large 4 pricing and AI video reality checks
Mistral Large 4 is in public preview at $0.68 per million input tokens, but its promised open weights aren't available yet. Kandinsky's video model is open to try, though five seconds of footage can take minutes to generate. We also look at Deutsche Telekom's AI savings targets and another model launch worth watching rather than adopting today.
The Big Story
Hacker News
Mistral Large 4
Mistral lists Mistral Large 4 as an open-weight multimodal model in public preview, with 49 billion active parameters, 1.05 trillion total parameters, a vision encoder, and a 1 million-token context window. Its API documentation lists the model as `mistral-large-4` and shows rates of $0.68 per million input tokens, $0.07 per million cached input tokens, and $2.09 per million output tokens. Despite the open-weight description, the Weights tab says “Coming soon,” and the page does not yet provide a license.
API users can compare its preview pricing and performance against their current model, but anyone planning self-hosting should wait for weights and license terms.
For Builders
AI TLDR
Kandinsky 6.0 Video — open MIT models make video with synced speech and sound
KandinskyLab has open-sourced Kandinsky 6.0 Video, with 3-billion-parameter Lite and 29-billion-parameter Pro models for generating five-second clips from text or images with synchronized 44 kHz audio and lip-sync. A separate super-resolution model can raise output to 1080p, and the repository lists ComfyUI and vLLM-omni support. Local use requires an NVIDIA GPU and Python 3.13 or 3.14; published non-distilled benchmarks range from 203 seconds for Lite HD on an H100 to 3,530 seconds for Pro Full HD on an RTX 5060 Ti, excluding model loading and encoding.
Try a sample clip or hosted demo before budgeting for local generation; five seconds of video can take several minutes or longer, depending on the GPU and model.
For Business
AI News
Deutsche Telekom targets €2.5 billion in savings from AI
Deutsche Telekom says it aims to cut indirect costs by ~€2.5 billion by 2030 versus 2023 through AI and automation, with a nearer-term target of ~€1.1 billion in gross savings outside the US by 2027. Its reported use cases include customer support, network management, software development, and administrative work; the company says its chatbot handled ~2.6 million service calls in the first half of 2026. It also plans an AI platform for small and midsize businesses to automate recurring customer-service and logistics tasks, but the article gives no launch date or pricing.
Treat the savings figures as company targets, not a guaranteed return for smaller firms; for now, measure one repetitive workflow’s cost and quality before committing to AI automation.
Worth Watching
TechCrunch AI
Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost
Reflection AI unveiled Beam, a text-only open-weight model with 501 billion total parameters, 23 billion active parameters, and a 1 million-token context window. The company claims it matches Z.ai’s GLM-5.2 on advanced reasoning benchmarks while using 3–4 times less inference compute, but those results haven’t been independently verified. Reflection says it will release the model weights and full technical details later in October, with distribution through cloud providers and open-source integrations.
Wait for the weights, licensing details, and independent tests before comparing Beam’s actual deployment cost with existing models.