Jump within the current audio brief. Story thumbnails mirror the cards below when available.
Same-day briefs for Sat Aug 15, sorted by release time. The selected date stays fixed.
Changing the date preserves the current brief type: AI News PM.
/ai-news-pm/2026-08-15.html.Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
Alibaba's Qwen team released Qwen 3.8-27B on August 14–15 under Apache 2.0 — a dense 27B multimodal model with 262K native context (extensible to 1M). Benchmark improvements are dramatic: DeepSWE 1.1 jumped from 13.3 to 42.2, OSWorld-Verified from 63.9 to 84.3, making it arguably the strongest locally-runnable multimodal model at the 30B scale.
Meta AI released Muse Glimmer on August 10, a 30B-parameter multimodal model under Apache 2.0 designed to run on a single consumer GPU. Built for always-on local agentic workflows, it supports 100+ languages with a 131K context window and is essentially the open version of Meta's closed Muse Spark flagship.
OpenAI's cybersecurity-specialized model autonomously discovered two previously unknown V8 vulnerabilities in Chrome that can be chained to corrupt memory and escape the V8 heap sandbox — before the model officially launched on August 10. Google patched both under CVE-2026-15903; security researchers are flagging the incident as proof of an alignment gap between specialized and standard models.
Gemini crossed 1 billion monthly active users on August 11, making it the fastest product in Google's history to reach that milestone. It climbed from 400 million users in May 2025 to clearing the billion mark in roughly 15 months. ChatGPT reportedly hit 1 billion in June, but that figure used weekly active users versus Gemini's monthly count — a comparison the press has mostly failed to scrutinize.
The Future of Life Institute's Summer 2026 AI Safety Index evaluated nine frontier AI labs across 37 indicators; the top score was Anthropic at C+ (2.66/4.0), with OpenAI and Google DeepMind at C, Meta at D+, and xAI, DeepSeek, and Mistral all receiving F grades. Seven independent reviewers ran the assessment across six domains including risk assessment, current harms, and existential safety — not a single company passed.
Google released HEIR (Homomorphic Encryption Intermediate Representation), an open-source MLIR-based compiler toolchain that converts pretrained models to run inference entirely on encrypted inputs — the server never sees the underlying data. Paired with Jaxite for GPU and TPU acceleration, it is the most practical step yet toward private AI inference without needing to trust a cloud provider.
OpenAI cut GPT-5.6 Luna from $1 to $0.20 per million input tokens on July 30, making it the free ChatGPT default with unlimited conversations. The company cited internal efficiency gains; analysts point to rising pressure from Chinese labs and Anthropic's aggressive model-ladder pricing as equally motivating factors.
xAI released Grok 4.6 on August 12, a post-training refinement of the 1.5T-parameter V9 foundation with better coding, reasoning, and instruction following. It now matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index at a score of 61 — two points behind Claude Opus 5 — while pricing holds flat at $2/$6 per million tokens with a 500K context window. A larger Grok 4.7 on a new 2.1T-parameter foundation is reportedly coming in late August.
Anthropic dropped the "introductory" label from Claude Sonnet 5's pricing on August 10, locking in $2 input and $10 output per million tokens as the permanent rate. Combined with Opus 5 launching last month at half the cost of the flagship Fable 5, Anthropic is building a stable, readable model ladder — a quiet but deliberate signal to enterprise buyers in a week full of price volatility.
DeepSeek — the Chinese lab celebrated all year for radical cost efficiency — raised its V4 Flash API pricing by 93% on August 14, from $0.14 to $0.27 per million tokens. The move is a counter-signal to the prevailing narrative about Chinese AI undercutting on price indefinitely, and suggests real infrastructure floors are emerging even for the most efficient labs.
