AI Product Updates Daily — May 27, 2026

Today's sweep covers Google's simultaneous rollout of Gemini Omni Flash and its forced migration away from the open-source Gemini CLI, Anthropic announcing a public release timeline for its cyberweapon-grade Mythos model, Microsoft shipping its May Copilot Studio bundle, Alibaba debuting Qwen Cloud at its first international conference, and a wave of API price cuts from Xiaomi and DeepSeek hitting today.

Google: Gemini Omni Flash goes live, Gemini CLI gets a deadline

Gemini Omni Flash is rolling out today to all Google AI Plus, Pro, and Ultra subscribers through the Gemini app and Google Flow, and at no cost to YouTube Shorts users through YouTube Create. 1

The model is the first in Google's new Omni family — a multimodal generation system that takes any combination of text, image, video, and audio as input and outputs video. Key capabilities include multi-turn conversational video editing (so each instruction builds on the last without breaking character or physics consistency), reference-driven generation from uploaded images or audio, and digital avatar creation using the user's own voice and likeness. All Omni-generated videos carry a SynthID watermark. API and enterprise access is scheduled for coming weeks.

Gemini Omni Flash rolling out across Google AI and YouTube Shorts — Gemini Omni brings multimodal video generation to all Google AI subscribers 1

On the developer tooling side, Google set a firm deprecation date: June 18 is when Google AI Pro, Ultra, and free users lose access to Gemini CLI, Gemini Code Assist IDE extensions, and Gemini Code Assist for GitHub. 2

The replacement is Antigravity CLI, a closed-source, server-side agent platform announced at Google I/O. Google describes it as "built for today's multi-agent reality," with a unified architecture that supports orchestrating multiple background agents and applies core improvements universally. Enterprise users and those with paid API keys keep Gemini CLI access.

The reception from developers has been rough. Reddit threads from Antigravity early adopters document token quotas expiring within six to seven prompts on the Pro plan — compared to users previously building entire projects with Gemini CLI while consuming only 13% of their quota. The open-source community is also unhappy: Gemini CLI had hundreds of active contributors; Antigravity's GitHub repository launched nearly empty. Google has acknowledged there won't be immediate feature parity.

Anthropic: Mythos gets a public release timeline

Anthropic published an update to Project Glasswing on May 26, announcing its intention to release a public version of Claude Mythos — the cybersecurity-specialized model capable of autonomously generating working exploits — once adequate safeguards exist. The company acknowledged those safeguards don't yet exist. 3

The near-term plan: expand Project Glasswing access from its current ~50 vetted partners (including Cisco, Oracle, and Microsoft) to U.S. and allied governments, then pursue broader public release. Anthropic's stated belief is that "Mythos-level models will become widely available in the next 6-12 months" across the industry regardless, and that building security tooling around the capability now is preferable to waiting. 4

Results from the Glasswing program so far: 10,000+ high/critical severity vulnerabilities identified across global software, of which 6,202 were in open-source projects. A third-party review of 1,752 high/critical findings confirmed 91% as valid, with two-thirds rated at the correct severity level. Among the discoveries was a certificate-forgery flaw in the wolfSSL cryptography library used by billions of devices. The flaw has been patched.

Anthropic released an accompanying toolset for vetted security researchers: custom instruction templates, a code-scanning suite with sub-agent orchestration, and a threat modeling tool that scans codebases to prioritize attack surfaces.

Microsoft Copilot Studio: May 2026 update bundle

Microsoft shipped several updates to Copilot Studio today, framed around closing the gaps in enterprise automation. 5

Computer-using agents reach general availability with new enterprise controls: secure credential management, per-automation model selection, and self-healing UI interaction that adapts when web interfaces or desktop layouts change. A new capability — embedding computer-use agents inside multi-step workflows — enters preview, letting teams combine API calls, approval gates, business logic, and adaptive UI interactions in the same automated system. Graebel, a global employee relocation firm, deployed this to process service orders across its proprietary Global Connect platform, which had no API.

The workflow designer gets a redesigned visual canvas (currently available in early-release environments) with an agent node concept: when a workflow step requires reasoning rather than deterministic if-then logic, teams can drop in an AI agent to handle that segment. Node-level testing is included for faster iteration.

On the platform layer: Work IQ gains a REST API and CLI for custom integrations, remote MCP server support, and agent-to-agent (A2A) communication — now generally available. The updated orchestration layer (auto-enabled in early-release) reports roughly 20% better evaluation performance and 50% lower token consumption compared to the prior generation.

Work IQ API architecture diagram showing how Microsoft 365 Copilot, Copilot Studio agents, and custom AI solutions connect to enterprise data — Work IQ API connects agents to Microsoft 365, Dataverse, and external systems through a unified layer 5

Real-time voice agents are generally available in North America through Dynamics 365 Contact Center, with server-to-server voice integration and full context handoff to human agents when escalations occur.

Copilot Studio computer-using agent setup interface — Copilot Studio's new GA computer-using agent interface 5

Alibaba Cloud: Qwen Cloud and Qwen3.7-Max at first international Qwen Conference

At its first international Qwen Conference in Singapore on May 26, Alibaba Cloud announced several products. 6

Qwen3.7-Max is now available on Model Studio in Singapore. According to Artificial Analysis's Intelligence Index, it ranks fifth globally and first among Chinese models with a score of 56.6, above Kimi-K2.6 and DeepSeek-v4-Pro-Max.

Qwen Cloud launches as a new AI-native cloud platform with three access modes: a Skills portal for agents (which converts 60+ cloud products into MCP-compatible function calls), a CLI for workflow integration, and a web interface for human users. The platform surfaces Alibaba's Qwen models, open-source models, and third-party offerings.

The JVS Agent Suite — including JVS Claw Teams (for cloud operations with 7×24 scheduling and centralized Skills distribution) and JVS Mobile (for autonomous task execution across mobile applications) — was also unveiled. Alibaba Cloud separately announced it joined the PyTorch Foundation as a Platinum member.

API pricing: MiMo-V2.5 cuts up to 99%, DeepSeek V4 Pro locks in 75% discount

Xiaomi MiMo-V2.5 API pricing was cut by up to 99% effective today (May 27, 00:00 CST), with no differentiation based on input length. The Token Plan billing system was also overhauled: existing subscriber quotas increase to 5–8× the previous levels, and all Credits for users currently within a valid plan period are fully reset. The cuts follow inference efficiency improvements including SWA-based KV Cache optimization via SGLang HiCache, which reduces data transfer between GPU memory, CPU memory, and SSD to roughly one-seventh of previous levels. 7

DeepSeek V4 Pro made its earlier 75% API price reduction permanent, according to coverage published May 26. The cut — initially described as temporary when first disclosed — is now a standing price floor. 8

Benchmarks: GPT-5.5 tops Google's Android Bench

Google's Android Bench leaderboard — updated May 18 — now ranks GPT-5.5 as the top AI model for Android app development, ahead of Gemini. 9 The prior update had Gemini 3.1 Pro and GPT-5.4 tied at the top. The leaderboard was also expanded to include open-weight models and new columns for latency, token consumption, and cost-per-run.

The benchmark evaluates models against real-world Android pull requests from public GitHub repositories — covering breaking-change resolution, wearable networking tasks, and Jetpack Compose migrations. Google VP of product Matthew McCullough said the goal is to "help model creators identify gaps and accelerate improvements." Andrew Filev, CEO of Zencoder, noted that while public benchmarks like Android Bench are useful for pushing model improvement, data contamination from training on public repositories can compress scores artificially — private evals on actual workloads often spread the rankings by a much wider margin.

Briefly noted

Anthropic Claude Opus 4.7 reached general availability approximately one week ago, with improvements in software engineering and long-horizon agentic tasks. Pricing is unchanged from Opus 4.6. It is currently the top-ranked model for Android coding in at least one comparative test, ahead of GPT-5.5, according to some third-party evaluators. 10
xAI Grok V9-Medium — Elon Musk confirmed the 1.5-trillion parameter base model has finished training and evaluation results are positive. Public release is expected within 2–3 weeks.
Microsoft Copilot Studio also shipped a separate update adding a natural-language migration assistant to Azure Copilot, designed to help organizations plan and execute complex infrastructure migrations through conversational input.
Pope Leo XIV published a 42,000-word encyclical on artificial intelligence (Magnifica Humanitas) on May 26, calling for international legal frameworks and independent oversight, and stating that lethal decision-making must never be delegated to AI. It is the first papal encyclical dedicated to AI. Anthropic researcher Chris Olah, speaking at the Vatican, noted that all frontier AI labs face incentives that can conflict with making the right choices. 11