Back to Blog

AI News Roundup: ChatGPT Enterprise on AWS, Meta's Muse Ember, Robot Manipulation Breakthrough & More — 9 November 2026

AI News
AI News Roundup: ChatGPT Enterprise on AWS, Meta's Muse Ember, Robot Manipulation Breakthrough & More — 9 November 2026

Today's AI news: ChatGPT Enterprise hits AWS Marketplace, Meta launches Muse Ember, AMP1 robots outperform specialist policies, and more.

••4 min read

Here is your daily digest of the most significant AI developments for 9 November 2026.

ChatGPT Enterprise Arrives on AWS Marketplace

OpenAI has made ChatGPT Enterprise available through the AWS Marketplace, allowing organisations to procure and bill for the product directly through their existing Amazon Web Services relationships. The move simplifies purchasing for enterprise customers who already manage significant AWS spend, removing the need for a separate commercial arrangement with OpenAI. It is a notable distribution milestone that puts ChatGPT Enterprise alongside thousands of other enterprise software tools available via AWS.

Meta Releases Muse Ember — Its Smallest, Most Efficient Model

Meta AI has published Muse Ember, described as the company's smallest and most efficient model to date. The release, surfaced via Meta's research blog, signals a continued industry push toward compact, cost-effective models that can run closer to the edge or at lower inference cost — a counterpoint to the ongoing race for ever-larger frontier models. Few technical details have been published at this stage, but the announcement has attracted considerable early attention.

Generalised AI Agents Outperform Specialist Robot Policies

A new paper on arXiv introduces agentic manipulation policies (AMPs) — an approach in which a generalist AI model directly commands a robot manipulator rather than relying on specialist vision-language-action models (VLAs) or world-action models (WAMs). The reference implementation, AMP1, achieves an 85.6% zero-shot success rate on the RoboLab-120 benchmark, compared with 49.2% for the leading VLA and 42.9% for the leading WAM. In real-world trials, AMP1 succeeded in 104 of 108 tasks across twelve scenarios and three robot types, reducing failure rates by up to 20 times relative to VLA baselines. The agent required no demonstrations or task-specific training, and developed its own strategies — including searching for hidden objects and using items as tools. When equipped with persistent memory, the system further reduced failure rates on bimanual manipulation tasks by 4.6 times compared with zero-shot deployment. The researchers argue that AMPs offer a substantially more general and robust foundation for robotics than specialist policies.

Agentic Reasoning Caching Cuts LLM Inference Costs by Up to 291×

Another arXiv paper introduces CPU-ARC (Agentic Reasoning Caching), a method that separates the reasoning phase of a task from its execution phase. Once a large language model has reasoned through a problem type, the resulting solution is compiled into a CPU-executable cache that can be applied to new instances of the same problem without re-invoking the full model. The paper reports cost reductions of 16 to 291 times compared with running the source frontier model, while maintaining on-par performance across difficult mathematical and scientific benchmarks. The caches are described as fully explainable, continually trainable, and robust to input variations — properties that are often difficult to achieve with standard LLM inference.

OpenAI API Cuts Engineering Search from an Hour to 4.5 Minutes

OpenAI has published a case study on Kuken Industries, a cooling tower manufacturer that used the OpenAI API to make 70 years of technical drawings searchable. Before the system, engineers spent close to an hour locating the correct drawing for a repair job. The AI-powered search reduces that to 4.5 minutes — a roughly 92% reduction in lookup time. The example is a concrete illustration of how AI document intelligence is delivering measurable productivity gains in traditional industrial settings.

Stanford's AI-Native Engineering Course Rethinks Curriculum for the Agent Era

Coverage from Latent Space highlights Stanford's CS146S course, led by Mihail Eric, which is grappling with a fundamental challenge: coding agents can now complete last year's assignments in a single prompt. The course's response is to have students engage with unfamiliar open-source repositories, propose meaningful changes, and work directly with open-source maintainers. The approach reflects a broader rethinking of what software engineering education should look like when AI can handle routine coding tasks — and what skills human engineers need to cultivate instead.

Frequently Asked Questions

What is ChatGPT Enterprise on AWS Marketplace?

ChatGPT Enterprise is now available through AWS Marketplace, allowing organisations to procure and bill for it via their existing Amazon Web Services purchasing relationships, simplifying the buying process for enterprise customers.

What is Meta's Muse Ember model?

Muse Ember is Meta AI's smallest and most efficient model, announced on 9 November 2026. It represents Meta's push toward compact, cost-effective AI models rather than purely scaling up to larger frontier models.

What did the AMP1 robot manipulation research find?

The AMP1 system, which uses a generalist AI model to control robot manipulators instead of specialist policies, achieved an 85.6% zero-shot success rate on the RoboLab-120 benchmark — far ahead of leading specialist approaches at 49.2% and 42.9%. In real-world trials it succeeded in 104 of 108 tasks with no demonstrations or task-specific training.

How does CPU-ARC reduce AI inference costs?

CPU-ARC (Agentic Reasoning Caching) works by compiling the reasoning an LLM performs on a problem type into a CPU-executable cache. New instances of the same problem can then be solved using the cache rather than the full model, reducing costs by 16 to 291 times while maintaining comparable performance on mathematical and scientific tasks.

Explore Topics

#ai-news#roundup