Skip to content
Daily BriefNewsDaily Brief

5 AI stories from October 15-18, 2026: Liquid AI's d1-ultra, Meta's deleted training clause, Reflection Beam on AWS, Perplexity's open-sourced sandbox

· by Pondero Newsdesk · 5 stories

AI news daily brief: 2026-10-18

Five stories from the past three days: a new decision-model benchmark claim, a legal walk-back at Meta, an open-weight coding model reaching a major cloud marketplace, an agent vendor open-sourcing its own security layer, and a lab's internal data on what AI agents are doing to its own headcount and calendars.

Liquid AI's d1-ultra claims an 8-point lead on the decision-model leaderboard

Liquid AI released d1-ultra on October 15, a multimodal decision model built for a narrow job: take text or images plus a fixed set of allowed answers, such as yes/no or a 0-1 score, and return a probability in one forward pass with no output tokens generated. Per Liquid AI's blog post, the model beats every rival it tested on the Artificial Analysis Decision Index by at least 8 points, which the company attributes to better-calibrated probabilities rather than overconfident guesses. It ships now through the Liquid API at $0.042 per million input tokens with no charge for output. That price point is not unique. Microsoft, OpenAI, and AWS each shipped their own decision models in the two weeks before d1-ultra, two of them at the identical per-token rate, turning classification-style AI workloads into a four-way price match before most buyers have heard the category name. Full story: Liquid AI's d1-ultra claims an 8-point benchmark lead as the fourth decision model to ship in two weeks.

Meta removed a clause from its commercial terms of use that had let the company "use any content" from a business account's connected domains to "provide, develop, improve, research, train, and maintain" its products. The deletion came two days after The Verge reported that legal experts flagged the wording as a possible route to scrape publisher sites for AI training, and TechCrunch confirmed the removal on October 17. Meta told reporters the language "only applied to our ads and business products" and that cutting it "does not change how Meta uses web crawlers." The company did not explain why wording it says never applied to crawling needed to be removed at all once lawyers started asking. Full story: Meta deletes AI-training clause from its commercial terms after legal pushback.

Reflection AI's 501B-parameter Beam model lands in SageMaker JumpStart

A 501-billion-parameter mixture-of-experts model with 23 billion active parameters is now deployable through Amazon SageMaker JumpStart in two US regions, AWS confirmed on October 16. Reflection AI's Beam, built for agentic coding, first previewed on October 5 under an Apache 2.0 license. Buyers get three paths in: a managed JumpStart endpoint, SageMaker HyperPod Inference, or a custom SGLang or vLLM setup on HyperPod. Full story: Reflection Beam on AWS SageMaker JumpStart: when self-hosting actually pencils out.

Perplexity open-sources SPACE, the sandbox behind its Computer agent

Perplexity announced on October 15 that it is open-sourcing SPACE, the AWS-Firecracker-microVM-based sandbox that has run every session of its Computer agent since July, Semafor reported. The code is released under Apache 2.0, and Perplexity lined up CoreWeave, Akamai, and Crusoe to offer it to other customers rather than keeping it as an internal moat. A named Perplexity engineer told Semafor that a red-team test pitting nine frontier models against the sandbox logged 118 attempts to escape it; none succeeded. Full story: Perplexity open sources the sandbox that has contained its Computer agent since July.

Anthropic's own data shows agents logging six times more hours than its staff

Agent work hours overtook human work hours at Anthropic company-wide in March 2026 and reached a six-to-one ratio by August, according to an internal report the company published on October 16. Junior-level technical hiring fell from 27% to 9% of new hires since early 2025, and TechCrunch's coverage of the report notes meeting time per employee rose 20% year over year. Anthropic attributes the meeting increase to human oversight becoming more important, not less, as agents absorb routine work. Full story: Anthropic says its AI agents now log six times more work hours than its employees.

Sources