Microsoft's First In-House Reasoning Model Headlines a Product-Heavy Build 2026
If you have been waiting to see whether Microsoft would build its own frontier intelligence rather than leaning entirely on OpenAI, Build 2026 is a clear answer. Mustafa Suleyman used the San Francisco keynote to unveil MAI-Thinking-1, the company's first dedicated reasoning model, alongside MAI-Image-2.5 and a faster MAI-Image-2.5-Flash for image generation. For anyone building on Copilot or the Azure AI stack, this matters because it gives Microsoft a route to ship reasoning and imaging without paying a margin to a third party, which tends to translate into lower prices and tighter integration over time.
The real test will come once independent benchmarks land. Treat the launch as a statement of direction as much as a finished product: the Copilot you use day to day may increasingly run on Microsoft's own models rather than GPT under the bonnet. The bar to clear on coding is Claude Code's 69.2 per cent on SWE-bench Pro.
Microsoft Shifts Copilot from Prompted Assistant to Autonomous Autopilot
The bigger shift at Build is conceptual: Microsoft is moving Copilot from something you prompt to something that acts on its own. It announced Autopilots (long-running autonomous agents), with the first named Scout, plus a forthcoming Copilot super app that folds the various Copilot surfaces into one. For everyday users this points to assistants that complete multi-step work across your apps without being nudged at each stage: a productivity leap or a trust question depending on your appetite for delegation. Developers get the plumbing too: a Windows Agent Framework, Azure Agent Mesh and an Agent Mode for Office 365.
Microsoft previewed Project Polaris, a homegrown coding model positioned as a future reasoning engine for GitHub Copilot. No benchmarks published yet; expect them alongside the Build session recordings.
Anthropic Widens Project Glasswing to 150 More Organisations Across 15-Plus Countries
If your organisation works in critical infrastructure, this is the access change to note. Anthropic expanded Project Glasswing, its preview programme for the cybersecurity-focused Claude Mythos model, to roughly 150 more organisations across more than 15 countries, deliberately pulling in sectors that were under-represented before: power, water, telecoms and healthcare. Anthropic says Mythos has already surfaced more than 10,000 vulnerabilities for existing partners in its first month. Access still depends on passing Anthropic's security vetting, so this is a controlled expansion rather than a general release.
For defenders, it is a signal that frontier models are moving from writing code to auditing it at scale, and a hint that a wider Mythos rollout may not be far behind. This was found on Anthropic's own newsroom before mainstream press coverage.
GPT-5.5, GPT-5.4 and Codex Reach General Availability on Amazon Bedrock, Including GovCloud
If you run workloads on AWS, you can now reach OpenAI's frontier models and its Codex coding agent without leaving Amazon's security, billing and governance perimeter. GPT-5.5, GPT-5.4 and Codex moved from limited preview to general availability on Amazon Bedrock, including GovCloud regions, which removes one of the main blockers to putting OpenAI models into regulated production.
For teams that were stuck choosing between OpenAI's models and their existing AWS compliance setup, that trade-off largely disappears. Codex landing natively on Bedrock is the standout, since it brings agentic code review and modernisation into the place many enterprises already build and ship.
Nvidia Uses COMPUTEX Week to Push AI onto the PC and into Robots
Nvidia detailed the RTX Spark superchip at GTC Taipei during COMPUTEX: a combined CPU and GPU part built with MediaTek for a new class of AI PCs from Dell, HP, Lenovo, ASUS, Microsoft Surface and MSI, aimed at running models locally without a round trip to a data centre. It also released a fresh batch of physical AI models as partners showed next-generation robots. For anyone weighing local inference, this dovetails neatly with Microsoft's Build push for on-device AI, so the two stories are best read together.
Industry Themes
Agentic AI moved from demo to default today. Microsoft's Autopilots and OpenAI's Codex-on-Bedrock both frame autonomous agents as the unit of work rather than chat, which means the practical question for builders is shifting from how to prompt a model to how much you are willing to let one run unattended. The model layer is also being verticalised by the platform owners: Microsoft shipping its own MAI reasoning and image models, plus the Polaris coding model, visibly reduces its dependence on OpenAI even as OpenAI deepens its reach inside AWS, a useful reminder that today's partners quietly hedge against one another.
Finally, local and physical AI gained ground, with Nvidia's AI PC chips and Microsoft's on-device push both pointing to inference leaving the data centre. If you care about latency, privacy or cost, this is worth tracking: the direction of travel is clear even if the products are still early. None of today's items surfaced as social-only exclusives; headline announcements from Build, Anthropic and the Bedrock launch all appeared on company channels simultaneously with press coverage.