AI in April 2026: agents go to work and a model too risky to release
April was the month the big AI providers stopped talking about agents as a future idea and started selling them as everyday business tools. It also brought a sharp reminder that more capable models bring new security questions, and a shift in how businesses can buy AI. Here are the developments that matter most, and what they mean for your organisation.
OpenAI launches GPT-5.5 and workspace agents
On 22 April, OpenAI introduced workspace agents in ChatGPT for its Business, Enterprise, Edu and Teachers plans. These agents run in the cloud, connect to a company's business tools and can work through multi-step tasks such as preparing reports or responding to messages. They can run on a schedule or in response to a trigger, can be deployed in Slack, and ask for approval before sensitive actions such as sending emails, creating calendar events or editing spreadsheets. Administrators get role-based access controls and visibility of how agents are configured and used.
The following day, OpenAI released GPT-5.5, which it describes as its strongest model yet for agentic work. OpenAI says it is designed to take "a messy, multi-part task" and "plan, use tools, check its work" through to completion, with particular strengths in coding, research, data analysis and creating documents and spreadsheets. It became available to paid ChatGPT users and in Codex straight away, with API access following.
What it means for your business: The combination matters more than either announcement on its own. Agents that can be scheduled, connected to your systems and made to ask permission before acting are much closer to how operations teams actually work. Start by mapping recurring, rules-based reporting and admin tasks, and decide in advance which actions should always need human sign-off.
Google turns Vertex AI into an agent platform
At Google Cloud Next '26 in late April, Google rebranded its Vertex AI platform as the Gemini Enterprise Agent Platform. It brings together tools to build agents (including a low-code Agent Studio), run long-running agents with persistent memory, and govern them through new features called Agent Identity, Agent Registry and Agent Gateway. Google also announced eighth-generation TPU chips, upgrades to its Gemini Enterprise app including Microsoft 365 interoperability, and new security agents for threat hunting and detection.
The theme throughout was moving AI from pilot projects into production. Google said nearly three-quarters of its cloud customers now use its AI products.
What it means for your business: Every major cloud provider now offers a full stack for building and managing agents. The deciding factors are less about which model is cleverest and more about governance: can you see which agents exist, what they have access to, and what they did? Build those questions into any platform evaluation.
Anthropic holds back its most capable model
On 7 April, Anthropic announced Project Glasswing, giving a limited group of organisations access to a new model, Claude Mythos Preview, for defensive cybersecurity work. Anthropic said the model had found thousands of previously unknown vulnerabilities, including in every major operating system and web browser, and that it "can surpass all but the most skilled humans" at finding and exploiting software flaws. Launch partners included AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft and Nvidia, and Anthropic committed $100 million in usage credits. Rather than release the model publicly, Anthropic said it would first develop new safeguards.
What it means for your business: AI is now finding software weaknesses faster than many teams can patch them, and attackers will eventually have similar tools. Keeping systems and third-party software up to date, and asking key suppliers how they are responding, has become more urgent. The same capability will also help defenders, so expect better automated security tooling to reach smaller firms over time.
OpenAI and Microsoft end their exclusive arrangement
On 27 April, Microsoft and OpenAI announced an amended partnership. Microsoft gave up its exclusive right to sell OpenAI's models, and in return stopped sharing revenue with OpenAI on products resold through its cloud, according to Bloomberg. Within days, OpenAI said its models and Codex would be available through other cloud providers, including Amazon Web Services via Amazon Bedrock, while Azure remains its primary cloud partner.
What it means for your business: If your organisation runs on AWS, OpenAI's models are now easier to use within your existing contracts, security set-up and data controls. More broadly, the main model providers are becoming available across the major clouds. That reduces lock-in and makes it more realistic to choose the best model for each task rather than the one your cloud happens to offer.
Practical takeaways
Identify two or three recurring reporting or admin tasks that a scheduled agent could take on, and pilot one with clear success measures.
Define which actions an agent may take on its own and which always need human approval, before you switch anything on.
Add agent governance questions (identity, permissions, audit logs) to any AI platform evaluation.
Review your patching routine and ask critical suppliers how they are responding to AI-driven vulnerability discovery.
Check which leading models are now available through your existing cloud provider before signing new contracts.
If you would like help working out where agents fit in your operations, or how to compare the growing range of platforms without getting lost in the detail, we would be glad to talk. Book a call with the Manta Global AI team and we can explore what is possible for your business.
Comments