AI Daily Brief — Curated by Claude
The Department of Defense finalized classified-network AI agreements with OpenAI, Google, Microsoft, Amazon Web Services, NVIDIA, SpaceX, and Reflection — granting access to Impact Level 6 and 7 environments used for mission planning, intelligence analysis, and weapons targeting. Anthropic was formally designated a "supply-chain risk" — a label historically reserved for adversary-linked firms — after the company refused to strip safety guardrails that would bar Claude from use in fully autonomous weapons or domestic mass surveillance. A federal judge blocked a broader ban last month, but a D.C. appeals court let the classified-network exclusion stand. In a twist, President Trump suggested last week that Anthropic was "shaping up," leaving the door open for a future deal.
siliconangle.com →OpenAI released GPT-5.5, billing it as its most capable model to date. The model matches GPT-5.4 per-token latency in real-world serving while delivering meaningfully higher intelligence — and uses fewer tokens to complete equivalent tasks. Capabilities include coding, computer use, research, data analysis, and cross-tool agent workflows. GPT-5.5 Pro is rolling out to Pro, Business, and Enterprise ChatGPT tiers, with the base GPT-5.5 available to Plus subscribers and via the API. OpenAI also published an updated system card with its strongest safeguards to date, including targeted red-teaming for advanced cybersecurity and biosecurity capabilities.
openai.com →xAI's Grok 4.3 Beta arrived with two headline capabilities: native video understanding — the model can now process and reason about video content in context — and document generation, letting Grok produce downloadable PDFs, formatted spreadsheets, and slide decks directly from a conversation. The update puts Grok closer to Gemini's recently expanded file-output suite and continues xAI's push toward a fully multimodal, productivity-oriented assistant. A Grok 5 roadmap remains the subject of community speculation, with no official release date announced.
docs.x.ai →Microsoft's new E7 "Frontier Suite" — bundling Microsoft 365 E5, Microsoft 365 Copilot, and the new Agent 365 ($15/user/month standalone) — became purchasable on May 1. The flagship feature is Cowork, which lets Copilot autonomously handle tasks across the M365 environment: drafting and sending emails, scheduling meetings, creating Word docs, posting in Teams, and managing calendar events on behalf of a user. Microsoft's multi-model strategy is also notable: the Researcher agent can use GPT to draft a response, then pass it to Claude for accuracy and citation review before delivery.
geekwire.com →OpenAI and Microsoft restructured their landmark 2019 partnership: Microsoft ends its Azure revenue-share arrangement with OpenAI, and in return OpenAI is free to distribute its models through Amazon Web Services and Google Cloud. The move dramatically broadens OpenAI's commercial reach ahead of what many expect to be an eventual IPO. For Microsoft, the trade gives it equity and IP access to future OpenAI models without bearing the same compute subsidy cost. Analysts called the restructuring a sign that the era of exclusive hyperscaler AI partnerships is fading as the frontier model market matures.
techcrunch.com →GitHub is moving its Copilot AI features to a consumption-based billing model, The Register reported. Instead of a flat per-seat subscription, usage of advanced Copilot capabilities will draw from a metered credit pool — a shift that mirrors broader industry moves toward pay-as-you-go AI pricing. The change is expected to lower costs for lighter users while letting heavy users scale up, and aligns GitHub's model with the token-level economics of the underlying foundation models it serves.
theregister.com →Cursor released a ground-up redesigned interface with Cursor 3, repositioning itself as an agent-first coding platform to compete head-on with Anthropic's Claude Code and OpenAI's Codex Desktop. The IDE remains Cursor's home advantage — it still ranks highest for inline autocomplete speed — but the new agentic layer lets it dispatch multi-step coding tasks that previously required switching to a CLI tool. A February 2026 Pragmatic Engineer survey of 906 engineers found Claude Code leading the field with a 46% "most loved" rating, making Cursor 3 an attempt to recapture ground lost to terminal-native tools.
thenewstack.io →Google I/O 2026, scheduled for May 19–20 in Mountain View, is shaping up to be one of the most AI-dense developer events the company has ever held. Expected: a Gemini 4 preview with sub-second multimodal reasoning, a full Project Astra integration enabling a single model that sees, hears, reasons, and responds in real time, and expanded Gemini Workspace features including file generation (Docs, Sheets, Slides, PDF, CSV, and more) now rolling out to users. Personalization features including Memories and chat import are also live in the UK, suggesting a broader rollout will be announced at the event.
blog.google →Reports emerged this week that OpenAI is developing a smartphone designed from the ground up around AI agents rather than traditional apps. The device would use a custom SoC co-developed with MediaTek and Qualcomm, with mass production targeted for 2028. Rather than launching apps, users would delegate tasks to AI agents that navigate software on their behalf — a concept reminiscent of earlier "app-free" AI device attempts, but backed by OpenAI's agentic model stack and distribution now spanning AWS and Google Cloud. No official confirmation has been issued.
openai.com/news →