AT&T swings 40% of its AI calls to open models to cut costs

AT&T is quietly remaking how 100,000 employees use AI, and the lesson for every company staring at a frontier-lab invoice is that open-source models are now the default, not the fallback. VP Mark Austin told reporters the telecom plans to keep its spend on Anthropic and OpenAI's closed models flat over the coming years, and the lever is routing more internal traffic to open and open-weight models like NVIDIA's Nemotron, Meta's Llama, and Google's Gemma.

The numbers are the story. Around 40 percent of employee AI calls now run on open models, a share AT&T aims to push to 60–70 percent over the next few years. Its Ask AT&T system alone processes roughly 45 billion tokens a day — code generation, HR lookup, sales call summaries, customer-support retrieval — an engine that, if fully hosted on closed frontier models, Austin estimates would cost over $100 million a year at published pricing. Austin says the open models he's tested match older Anthropic and OpenAI releases "or do better," and he puts the gap between open weights and the frontier at 6–10 months and shrinking.

Cost routing is doing the heavy lifting. AT&T deploys model-routing that assigns each request to the cheapest model that can do the job — summaries go to open models, hard codegen still gets the top closed models — and after plugging in the LiteLLM routing layer, it cut the cost of advanced coding tasks by up to 56 percent while output quality fell just two percent. The company also runs some open-model workloads on its own data centers with NVIDIA and AMD silicon rather than renting cloud GPUs. It's separately evaluating DeepSeek and Moonshot as they weigh China's open-source models, but hasn't put either into production yet.

This matters well beyond one carrier. It's a public, dollar-signed vote that availability of cheap open models is eroding the pricing power OpenAI and Anthropic bank on right as they slug it out in a very public price war — OpenAI just cut GPT-5.6 Sol API pricing by over 20 percent. We flagged that fight this morning in OpenAI's price war gambit is a warning shot at Anthropic's IPO. AT&T's routing math shows why the pressure is structural, not just competitive: when your biggest customers can swap a modular open model in for a frontier API and lose two percent quality, the closed-model premium has to keep earning its keep.

If you're an AI buying team reading this: have you benchmarked your own token mix against open models yet? Tell us in the comments.

AI 101 — How do AI video generators work?

An AI video generator is a model that starts with a grid of random noise and repeatedly refines it into a moving picture that matches your prompt — the same trick image generators use, stretched across time. Nothing is filmed, and nothing is stitched together from clips the model has seen. The pixels in the finished video were invented, frame by frame, all at once. Why it matters right now Kling 4.0, Kuaishou's flagship video model, went fully live on October 15 with up to 30 seconds of vid

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Read more

AI 101 — How do AI video generators work?

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone