What Google, Anthropic, OpenAI and Microsoft Released in May 2026, and the EU's AI Act Delay

Title: What Google, Anthropic, OpenAI and Microsoft Released in May 2026, and the EU's AI Act Delay

Date: 5 June 2026

Type: Blog

Author: SAASiQ (contact@saasiq.ai)

Word count: 1539 words

Reading time: 6 min

Published: 05-06-2026

Tags: #AI #ThoughtLeadership #EnterpriseAI #Governance #AIAgents #Workforce #Innovation #GoogleIO #OpenAI #Anthropic

Google, Anthropic, OpenAI, xAI and Microsoft all released new models or agent tools in May, most of them designed to carry out multi-step work with less supervision than a chat assistant needs. In the same month EU negotiators agreed to postpone the AI Act's high-risk rules, and Oracle's 26B update reached production with its new agentic applications switched off by default. A Bipartisan Policy Center analysis of Lightcast data put the growth in US job postings asking for AI skills at 144 per cent in a year.

Google I/O: Gemini 3.5 Flash and Antigravity 2.0

Google opened I/O on 19 May with Gemini 3.5 Flash, generally available the same day in the Gemini app, the Gemini API and Gemini Enterprise. Google says it outperforms Gemini 3.1 Pro on coding and agentic benchmarks, including 76.2 per cent on Terminal-Bench 2.1, and Sundar Pichai said in his keynote that it runs four times faster than other frontier models at less than half the price of comparable ones. Google said Gemini 3.5 Pro would follow in June.

Antigravity 2.0, Google's agent development platform, is now a standalone desktop application in which a main agent breaks a task down and hands the pieces to subagents running in parallel, with a command-line tool and an SDK in preview alongside it. Google Cloud customers get it through the Gemini Enterprise Agent Platform, and a new $100 a month AI Ultra plan gives five times the Antigravity capacity of the Pro plan. Google's own write-up says only that its researchers had agents build a functional operating system. The figures from the keynote demonstration, as reported by AlphaSignal on 27 May, were 93 subagents working for about 12 hours, more than 15,000 model requests and 2.6 billion tokens, at a compute cost of under $1,000.

Pichai also gave usage numbers. Google now processes more than 3.2 quadrillion tokens a month, against about 480 trillion a year earlier, the Gemini app has more than 900 million monthly users (400 million a year ago), and more than 375 Google Cloud customers each process over a trillion tokens a year. Gemini Spark, a personal agent that runs around the clock on a cloud virtual machine and connects to third-party tools through the Model Context Protocol, began rolling out to trusted testers. Google also announced Gemini Omni, which generates output from any input, starting with video, and began rolling out Gemini Omni Flash to AI Plus, Pro and Ultra subscribers.

Anthropic: Claude Opus 4.8 and dynamic workflows

Anthropic released Claude Opus 4.8 on 28 May at the same price as Opus 4.7, $5 per million input tokens and $25 per million output tokens. Fast mode, which runs the model at 2.5 times the speed, now costs $10 and $50, a third of what it cost on earlier models. Users of claude.ai and Cowork can now choose how much effort the model puts into a task, with high as the default.

Dynamic workflows, a research preview in Claude Code on the Enterprise, Team and Max plans, lets one session run hundreds of subagents in parallel. Anthropic aims it at jobs such as codebase migrations running to hundreds of thousands of lines. It says Opus 4.8 is about four times less likely than its predecessors to let a flaw in code pass unremarked, and that it scored 84 per cent on Online-Mind2Web, a test of browser agents. Anthropic also said its Mythos-class models, in preview for cybersecurity work, would become generally available in the coming weeks once the safeguards were ready.

The same day Anthropic closed a $65 billion Series H at a post-money valuation of $965 billion, and on 1 June it confidentially submitted a draft registration statement for an IPO to the US Securities and Exchange Commission.

OpenAI and xAI: coding agents on the desktop

OpenAI's model release came in April, when GPT-5.5 launched on 23 April. Its May releases were about where Codex, its coding agent, can work. On 7 May it launched Codex for Chrome, an extension that lets Codex test web apps, pull context from several tabs and use the browser's developer tools without taking the browser away from the user. OpenAI said Codex had more than 4 million weekly active users, eight times the number at the start of the year.

On 29 May OpenAI extended computer use in the Codex app to Windows 11, so Codex can operate desktop applications by seeing the screen, clicking and typing. On Windows it runs in the foreground and takes over the active desktop for the length of the task, whereas the macOS version can work in the background. Tech Times reported on 1 June that the feature was not available at launch in the European Economic Area, the UK or Switzerland, and that OpenAI had given no date for those regions.

xAI, part of SpaceX since February, put Grok Build 0.1 into public beta on its API on 29 May. It is the model behind xAI's Grok Build coding agent, priced at $1 per million input tokens and $2 per million output tokens, runs at more than 100 tokens a second and is also offered through OpenRouter and Vercel AI Gateway.

Microsoft: controls for agents, then its own models

Microsoft made Agent 365 generally available on 1 May at $15 per user per month, and it is also included in Microsoft 365 E7. Microsoft describes it as a control plane to observe, govern and secure agents. It keeps a registry of agents, which can now sync with AWS Bedrock and Google Cloud in preview, and it uses Defender and Intune to find agents running locally on devices, OpenClaw among them.

In Copilot Studio, computer-using agents became generally available in May across all commercial Power Platform geographies. They work with websites and desktop applications through the user interface, which is aimed at processes where no API exists, and come with managed credentials and a choice of model. Agent-to-agent communication in Copilot Studio also became generally available.

At Build on 2 June Microsoft released seven models under its MAI name, trained without distillation from OpenAI or any other outside model. One of them, the speech recognition model MAI-Transcribe-1.5, is already built into Teams.

Oracle 26B: off until an administrator switches it on

Oracle's 26B quarterly update went to test environments on 1 May and production on 15 May, and it documents several of the agentic applications Oracle announced in March. They do not appear on the day. The Cost Accounting Close Workspace, for example, stays off until an administrator sets the profile option ORA_CST_PERIOD_CLOSE_AGENTIC_APP_ENABLED to Yes at site level.

To use agents on a Fusion page, a user's configured job role also has to contain the Fai Genai Agent Runtime Duty, with permission groups enabled on that role in the Security Console. An agent works within the access its user already has, so the role design done for a Fusion implementation is also the security design for its agents.

The EU postpones its high-risk rules

On 7 May negotiators for the Council of the EU and the European Parliament reached a provisional agreement on the Digital Omnibus on AI. Obligations for stand-alone high-risk systems under Annex III, which covers uses such as employment, education, biometrics and critical infrastructure, had been due to apply from 2 August 2026 and now apply from 2 December 2027. High-risk AI embedded in regulated products under Annex I moves to 2 August 2028.

Generative systems placed on the market before 2 August 2026 get until 2 December 2026 to meet the rules on marking synthetic content. The AI literacy duty is softened from ensuring a sufficient level of literacy to taking measures to support it. The Parliament and Council still have to adopt the text formally, which was expected before 2 August. UK organisations are outside the Act's direct reach unless they operate in the EU.

Hiring and use at work

The Bipartisan Policy Center's April analysis of Lightcast job-posting data found that US postings requiring AI skills had grown 144 per cent in a year, while postings overall grew 7 per cent over the same period.

The CNBC and SurveyMonkey quarterly AI and jobs survey, published on 14 May and based on 3,597 workers and students polled from 17 to 21 April, found that 65 per cent of workers had at some point avoided using AI because of moral, environmental, privacy, accuracy or other concerns. Among workers, 28 per cent cited moral or ethical concerns and 19 per cent environmental ones. Of those who use AI daily or weekly, 73 per cent said it made them more productive and 68 per cent said it saved them time.

Most of the enterprise tools described here reach staff through a setting that an administrator or developer controls: a profile option in Fusion, a licence and registry in Agent 365, an extension or desktop permission for Codex. SAASiQ's view is that an inventory of which of these are enabled, and for whom, is a more useful first step than evaluating each new model as it arrives.

Oracle reports its fiscal 2026 fourth-quarter results on 10 June, Google has said Gemini 3.5 Pro is due this month, and the Omnibus goes to the Parliament and Council for formal adoption.

SAASiQ - Intelligent Solutions for SaaS ©