← All updates

Hugging Face Opens Carbon-X DNA Model

Plus: OpenAI and Anthropic Slash API Prices, Google and Meta Opens TorchTPU

Good Morning, Model Builders!

Frontier API prices just fell by up to 80%, models are learning to read DNA, and TPUs now speak native PyTorch. The AI stack is getting cheaper and more open at every layer.


Hugging Face Opens Carbon-X DNA Model

Photo: Warren Umoh, Unsplash

TL;DR: Hugging Face released Carbon-X, a 14.8 billion parameter open DNA language model built on Qwen3-14B. It reads raw DNA inside a chat, answers in plain English and averages 66.7% across five DNA benchmarks. The weights ship under Apache 2.0, so teams can fine-tune and self-host it freely. Read more →


OpenAI and Anthropic Slash API Prices

Image: Max Woolf's Blog

TL;DR: A new analysis tracks how OpenAI and Anthropic cut API prices by up to 80% since June. GPT-5.6 Luna fell 80%, Claude Opus 5.5 costs 40-60% less in total than Opus 5, and the GPT-6.1 Sol family is about 75% cheaper, largely through lower cache read prices that dominate agent workloads. Read more →


Kimi Resets Quotas After Rate Limits

Photo: Taylor Vick, Unsplash

TL;DR: Moonshot apologized for rate limits and wasted quota on the K3.1 Pro subscription plan. Every subscriber on every plan gets a one-time reset restoring 25% of monthly quota, about one week of usage, and the plan now trims unnecessary subagent calls and output length so quota lasts longer. Read more →


Nanbeige4.3-33B Targets Small Agent Models

Photo: Fernando Andrade, Unsplash

TL;DR: The Nanbeige team released Nanbeige4.3-33B, claiming state-of-the-art agentic results among open models near 30 billion parameters, with strong agentic coding scores and cost-efficient inference inside mainstream agent harnesses. It will also serve as the teacher for smaller Nanbeige models, starting with Nanbeige4.3-3B. Read more →


OpenAI Resumes Frontier Training Offline

Image: theregister

TL;DR: OpenAI cut internet access during model training and resumed training its most capable models on October 1. The 11-day pause followed a research model reaching the internet through a configuration error; during it, AI agents red-teaming OpenAI's systems found a Brotli vulnerability, since patched, and gaps in its safeguards, a reminder that safety gates can shift frontier release timelines. Read more →


Google and Meta Open-Source TorchTPU

Image: InfoQ

TL;DR: Google and Meta open-sourced TorchTPU, a native PyTorch backend for Google TPUs that replaces PyTorch/XLA. Porting a model can be as simple as switching to("cuda") to to("tpu"), and Meta plans to move larger LLM workloads onto TPUs with it, widening hardware choice beyond Nvidia GPUs. Read more →


Google Open-Sources ML Drift GPU Engine

Image: InfoQ

TL;DR: Google open-sourced ML Drift, the GPU engine behind LiteRT, under Apache 2.0. It runs on-device inference across OpenGL ES, OpenCL, Metal and WebGPU, adds tuning for Arm, Intel and Qualcomm GPUs, and extends to Windows and Linux, making local generative AI easier to ship across mixed hardware. 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.