← All updates

OpenAI Opens Decisions API to All

Plus: Cohere Ships Rerank 5, Kling Makes Long Video Behave

Good Morning, Model Builders!

OpenAI opens its Decisions API to every customer, and Cohere ships Rerank 5 at $0.05 per million tokens.


OpenAI Opens Decisions API to All

Image: Neowin

TL;DR: OpenAI made its Decisions API generally available, a reconfigured ChatGPT Luna tuned for fast, cheap choices with confidence scores. Latency drops from one to three seconds to roughly 100 to 180 milliseconds, and the public version adds voice and visual inputs. It targets the decision-model niche TypeSafe opened with Jev. Read more →


Cohere Ships Rerank 5 With 32K Context

Image: Cohere

TL;DR: Cohere released Rerank 5 in Pro and Fast tiers with a 32K context window, passage-level Context Focus extraction and calibrated relevance scores. The Pro tier scores 82.2 on the BRIGHT retrieval benchmark and costs $0.05 per million tokens, a cheap upgrade for teams building retrieval pipelines in front of large models. Read more →


Kling Makes Long Video Behave

Image: Kling AI

TL;DR: Kling AI rolled Kling 4.0 into OnSolo, OpenArt and Luma, giving creators a more controlled video model for longer shots. The release supports 30-second native clips, up to 15 reference images and 10 keyframes per job, plus 21:9 output, steadier fast motion and tighter lip sync. It aims to keep characters, camera movement and story continuity intact across one generated sequence. Read more →


Nvidia Extends Confidential Computing to Bare Metal

Image: NVIDIA Blog

TL;DR: Nvidia added Confidential Bare Metal to its confidential computing stack, letting frontier model builders run their own software on dedicated third-party cloud, neocloud and enterprise hardware while keeping control of model weights and data. The protections now extend across Grace Blackwell and Rubin systems, alongside container and virtual machine options. Read more →


Samsung Electro-Mechanics to Supply Nvidia Substrates

Photo: Umberto, Unsplash

TL;DR: Samsung Electro-Mechanics announced a partnership to supply Nvidia with ABF package substrates for AI processors and networking switches. Substrates are one of the packaging inputs that cap how many accelerators reach data centers, so a second large supplier adds headroom to the GPU pipeline that model training and inference depend on. Read more →


IBM Gives Open Speech a Real Upgrade

Image: Hugging Face

TL;DR: IBM released Granite Speech 5.0, led by a new Apache 2.0 2B speech-language model for multilingual transcription and translation. Trained on 390,000 hours of public audio, it adds 2,000-term keyword biasing and stronger performance in noisy rooms. IBM also refreshed its 470M Turbo CTC model and 4.1 Plus model for speed, multilingual use and speaker attribution. Read more →


Diagnosis Gets an AI Care Team

Image: Microsoft AI

TL;DR: Microsoft AI introduced MAI-DxO, a diagnostic orchestrator that has multiple AI doctors debate hypotheses, request tests and revise their views as evidence arrives. In its new SDBench evaluation, Microsoft reports 83% accuracy on difficult NEJM cases and 80% on emergency-department cases while choosing cheaper tests. The company is open-sourcing parts of the benchmark, not the clinical system itself. Read more →


Enterprise Tables Get Their Own Models

Image: NVIDIA

TL;DR: Nvidia released Kumo Structured, an open family of foundation models built for relational and tabular business data rather than chat. Developers can use pretrained models to forecast outcomes, score fraud risk, detect anomalies and recommend actions, then tune them on private data. Nvidia pairs the weights with Kumo Engine, an open stack for deploying them without custom pipelines. Read more →


Top AI models. Lower costs.

Access GPT, Claude, Gemini, and more through one OpenAI-compatible API—and save up to 70% on AI API costs. Create your DIT account, get your API key, and start building.