The week AI alliances redrew the map
of enterprise infrastructure

The most strategically significant Gemini development this week was not a software release; it was a deployment at physical scale. General Motors announced that Gemini is rolling out to approximately 4 million model-year 2022 and newer vehicles, covering Chevrolet, Cadillac, Buick, and GMC models equipped with Google built-in. The update replaces Google Assistant with a conversational AI that can handle follow-up questions, draft and translate text messages hands-free, manage in-car entertainment from natural language, and understand context across a conversation without requiring a restart. The rollout is staggered over several months and requires an active OnStar connection, a signed-in Google account, and user opt-in. GM has indicated a more deeply integrated, OnStar-informed AI assistant is planned later this year. On the product side, Google launched Gemini 3.1 Flash TTS Preview, a cost-efficient text-to-speech model, released new Gemma 4 open weights in two sizes via AI Studio and the Gemini API, and introduced Flex and Priority inference tiers to give developers more control over the cost-versus-latency trade-off.

Google also pushed Gemini further into living rooms and smart home environments this week. Gemini features arrived on Google TV via a dedicated Create tab, allowing users on TCL devices to generate and remix images and video clips using Nano Banana and Veo directly from their television. The April Gemini Drops update confirmed the Gemini app is now natively available on Mac, bringing it into direct desktop competition with Claude and ChatGPT. Personal Intelligence, the feature connecting Gemini to a user's Google apps for personalised responses, began its international rollout. Meanwhile, Gemini for Home received a material speed improvement: controlling smart lights and plugs is now up to 1.5 seconds faster, and alarm and timer commands execute near-instantly. Leaked code in the Gemini beta also suggests a forthcoming Proactive Assistance capability that would allow Gemini to monitor selected apps, notifications, and screen context to generate real-time suggestions without being prompted, a development that carries both productivity promise and privacy implications for enterprise IT teams.

The most consequential OpenAI development of the week was not a model release; it was a change in commercial access. On April 27, Microsoft's exclusivity agreement with OpenAI concluded. The following day, OpenAI and AWS announced a major expanded partnership bringing three capabilities to Amazon Bedrock in limited preview: OpenAI frontier models including GPT-5.5, the Codex coding agent, and Amazon Bedrock Managed Agents powered by OpenAI. The deal is backed by a $50 billion Amazon investment and grants AWS exclusive third-party cloud distribution rights for OpenAI Frontier, the company's enterprise agent platform. For AWS customers, this means OpenAI models now operate within the full suite of enterprise controls they already depend on, including IAM-based access, PrivateLink connectivity, encryption, CloudTrail logging, and the ability to apply OpenAI usage toward existing AWS cloud spending commitments. The significance here is not technical; it is procurement and governance. Organisations that previously could not adopt OpenAI at scale because it sat outside their AWS infrastructure framework now have a clear path.

GPT-5.5 itself launched to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex earlier in the week. OpenAI describes it as its most capable model for agentic and professional work, with particular gains in multi-step reasoning, coding, document creation, and the ability to navigate complex tasks with minimal human direction. A GPT-5.5 Pro variant is available for Pro, Business, and Enterprise plans. Separately, OpenAI introduced Advanced Account Security, an opt-in setting for personal accounts that replaces password-based login with phishing-resistant passkeys, adds shorter session windows, login notifications, and recovery key controls. For organisations with security-conscious employees handling sensitive work in ChatGPT, this feature is worth activating immediately. The ChatGPT Business plan also launched Workspace Agents this week, allowing organisations to build and deploy agents for repeatable tasks, connect them to Google Drive, SharePoint, Slack, and other tools, schedule recurring runs, and share agents across the workspace directory.

Microsoft marked May 1 with the general availability of Microsoft 365 E7, its new Frontier Suite, and Agent 365, a companion product designed to make agentic AI the organising layer of enterprise work. The E7 tier represents Microsoft's clearest commercial signal yet that AI agents are no longer an optional add-on but a core part of the M365 product line. Agent 365 enables organisations to build, deploy, share, and schedule agents that work across email, meetings, documents, and workflows with user oversight built in. The April What's New in Microsoft 365 Copilot roundup also confirmed several material updates to core applications. Excel now has a Chat and Edit switcher, a plan mode for step-by-step reasoning, and the ability to invoke Python directly from Copilot prompts for data analysis, with visual output capability. PowerPoint allows users to select which image model to use for slide generation, including GPT-Image, Flux, or an Auto setting. Copilot in Outlook has been updated to write first drafts directly in the email canvas and iterate with the user through clarifying questions, reducing the copy-paste friction that has frustrated many users since Copilot's initial launch.

A detail worth noting for enterprise IT leaders: Excel has added support for both GPT-5.5 and Claude Opus 4.7 as selectable models within Copilot, reflecting Microsoft's broader strategy of offering model choice within the M365 environment rather than locking users to a single AI provider. This is a significant development for organisations evaluating which AI works best for specific task types within their existing Microsoft workflow. The mobile Copilot app received a redesign this week with a chat-first layout, liquid glass styling, text formatting in prompts, and improved citation display. Copilot Notebooks launched a wave of new capabilities in preview, including the ability to generate Word documents and PowerPoint presentations directly from notebook content, reference SharePoint and OneNote, and share notebooks with M365 Groups. Dragon Copilot, Microsoft's AI tool for healthcare documentation, also reduced its per-user list price from May 1, widening access across clinical environments.

Anthropic's most significant product development this week was for Claude Code, its developer-focused autonomous coding tool. Claude Code received two notable capability updates. First, Routines launched in research preview for Pro, Max, Team, and Enterprise users with Claude Code web enabled. A Routine bundles a prompt, a code repository, and relevant connectors into a single configuration that can run on a schedule, trigger from an API call, or fire automatically from a GitHub event such as a new pull request. Crucially, Routines execute on Claude's web infrastructure rather than a local machine, meaning they can run continuously without requiring a developer's computer to be active. This represents a shift from Claude Code as a tool developers use interactively, to Claude Code as a persistent automated colleague that operates around the clock. Anthropic has built in daily run caps that scale by plan tier. Second, Claude Code on the desktop app received an ultrareview automation subcommand, allowing code review to be triggered non-interactively from CI pipelines and scripts, and Windows users no longer need Git for Windows installed, as Claude Code now falls back to PowerShell.

On the consumer and creative professional side, Anthropic launched eight new connectors this week aimed at creative workflows: Ableton, Adobe, Affinity, Autodesk Fusion, Blender, Resolume Arena and Wire, SketchUp, and Splice. These integrations allow Claude to access and work within professional creative tools, opening the platform to a new audience of designers, architects, musicians, and video producers beyond its established base of knowledge workers and developers. The Claude platform API also formalised the retirement of the 1 million token context window beta for Claude Sonnet 4.5 and Sonnet 4, directing users to migrate to Sonnet 4.6 or Opus 4.6, where the 1 million token context window is generally available at standard pricing with no beta header required. For enterprise teams running long-context workloads, such as document review, contract analysis, or large codebase processing, this is a straightforward migration that removes the experimental designation from a capability they may already depend on.

Perplexity's most strategically relevant development this week was the enterprise launch of the Comet browser. Comet, Perplexity's AI-native browser, is now available for enterprise organisations with the ability for IT administrators to deploy it silently across macOS and Windows devices via Mobile Device Management tools. Enterprise administrators can configure hundreds of browser policies and control exactly which actions the AI agent embedded in Comet can take on behalf of employees. The Comet Assistant handles in-page research, content summarisation, and autonomous multi-step tasks including booking, email management, and form completion. The move from consumer download to managed enterprise deployment is a meaningful step; it positions Comet not as a personal productivity experiment but as an IT-administrable business tool that can be governed and standardised across a workforce. For leaders who have been watching agentic AI with interest but have been concerned about ungoverned employee adoption, Comet Enterprise addresses that concern directly.

Perplexity also opened the waitlist for Personal Computer, an always-on AI agent that runs on a dedicated Mac mini in the user's environment. Personal Computer monitors triggers, executes proactive tasks, and carries work forward around the clock, with every sensitive action requiring explicit user approval and a full audit trail available. While Personal Computer is a consumer-facing product for now, the capability it demonstrates is directly relevant to enterprise leaders thinking about what a persistent AI-backed research and operations function could look like. Separately, Perplexity extended memory to work within its Model Council feature, which allows a single query to be processed simultaneously across multiple AI models. The practical effect is that Perplexity can now pull relevant personal context into comparative model responses, making Model Council more useful for personalised decision-support queries. Perplexity's annualised recurring revenue crossed $450 million in March 2026, confirming it is one of the fastest-growing AI platforms by commercial velocity.

The defining story of the week in AI geopolitics was China's National Development and Reform Commission formally blocking Meta's $2 billion acquisition of Manus, the Singapore-based autonomous agent startup, on April 27, 2026. The ruling ordered both parties to unwind the deal, despite the fact that Meta had already integrated Manus technology into its AI operations, approximately 100 Manus employees had joined Meta's Singapore offices, and backers including Tencent and HongShan had already received their proceeds. The Chinese government's stated concern centred on Manus's origins: its core research and development took place in China, its data was generated there, and its founding talent, including two co-founders who are currently subject to a travel ban, remain tied to the country. Beijing's position, articulated in Chinese state media, is that regardless of where a company is registered, if its technology and talent originated in China, cross-border transfers require regulatory approval and can be blocked on national security grounds. This ruling applies even to companies that have undertaken what analysts are now calling "Singapore washing," the practice of relocating corporate headquarters to neutral jurisdictions to reduce the appearance of Chinese ties.

The practical consequences for Meta are significant. The company had positioned Manus's agentic AI capabilities, which allow autonomous web navigation, file management, and multi-step task execution, as central to its Superintelligence Labs strategy. The forced unwind raises questions about what technology Meta can retain, how integrated Manus's systems have become with Meta's own infrastructure, and whether the approximately 100 employees who joined from Manus will remain. Meta is publicly stating the transaction complied with applicable law and expects an appropriate resolution, but analysts note Beijing has limited leverage concern: Meta's social platforms are blocked in China, and the company generates essentially no revenue from the country. For the broader business world, analysts have drawn a sharp conclusion: any technology company founded in China, regardless of subsequent relocation, faces a materially higher regulatory risk profile for cross-border M&A. This is now a live due diligence requirement, not a theoretical one.

xAI's most commercially significant move this week was the general availability of two standalone audio APIs: Grok Speech to Text and Grok Text to Speech. Both are built on the same voice infrastructure that powers Grok Voice, Tesla in-vehicle systems, and Starlink customer support operations. The Speech to Text API supports transcription in 25 languages with both batch and streaming modes, speaker separation across multichannel audio, and advanced formatting that converts spoken language into properly structured text output, including normalising numbers, dates, and entity names. Pricing is set at $0.10 per hour for batch processing and $0.20 per hour for streaming. The Text to Speech API generates natural-sounding speech with expressive voice tags that allow control over tone and emotion. The voice agent API, which allows developers to build full voice-driven AI applications, also gained a new model variant this week: grok-voice-think-fast-1.0, which is now accessible through the xAI Voice Agent documentation. Separately, xAI expanded its Batch API to support image generation, image editing, and video generation alongside chat completions, broadening the range of automated content workflows that can be run asynchronously at scale.

The context for these developments is that xAI, now owned by SpaceX following the February 2026 acquisition, is executing against a clear commercial road map: convert Grok from a consumer chatbot visible on X into an enterprise-grade AI API platform that competes directly with OpenAI, Anthropic, and Google on developer workloads. The voice API launch is a direct challenge to ElevenLabs, Deepgram, and other specialised audio AI providers, positioning Grok's voice capabilities as part of a broader, integrated stack rather than a standalone product. For business leaders who use X daily, the practical news this week is that Grok 4.1 remains the current available model, with Grok 5 still expected in Q2 2026 after missing its original Q1 target. The Colossus 2 supercluster in Memphis is confirmed operational and actively training the next generation model, which has been described as a 6 trillion parameter architecture. Grok's enterprise API also now includes a cost tracking field on every API response, giving finance and operations teams precise, per-request visibility into AI spend without requiring separate usage dashboards.

Mistral launched two significant capabilities on April 29. The first is Mistral Medium 3.5, a 128 billion parameter dense model with a 256,000-token context window, released in public preview. The model is available through Mistral Vibe CLI, Le Chat, and the API for Pro, Team, and Enterprise users. Its headline capability is long-horizon autonomous coding: Mistral has deployed remote coding agents that run in isolated cloud sandboxes, complete extended sessions independently, and notify users on completion. Sessions can be started from the web interface or the terminal and can be moved between local and cloud environments while preserving state and any required approvals. The model integrates instruction-following, reasoning, and coding in a single architecture with variable image support through a newly trained vision encoder, and has been benchmarked at 77.6% on SWE-Bench Verified, a measure of autonomous software engineering task completion. The open weights are released under a modified MIT licence and the model can be run with as few as four GPUs, making self-hosted enterprise deployment genuinely feasible for organisations with existing GPU infrastructure.

The second launch is Work Mode for Le Chat, Mistral's consumer and enterprise chat interface. Work Mode is an agentic capability powered by Mistral Medium 3.5 that allows Le Chat to read and write across multiple tools simultaneously: email, calendar, web, and connected documents. The agent handles research and productivity tasks and surfaces actions for user approval before executing them, which addresses the governance concern that makes many enterprise leaders reluctant to deploy fully autonomous agents. For European organisations specifically, this combination of open weights, self-hostable infrastructure, on-premises deployment flexibility, and a capable agentic chat interface positions Mistral as a credible alternative to OpenAI and Anthropic for leaders who need to keep data within European jurisdiction. Mistral also separately released Workflows in public preview this week, an orchestration layer for enterprise AI processes built on Temporal's durable execution engine. Workflows allows organisations to build AI-powered processes in Python with human-in-the-loop approvals, traceable execution, and deployment flexibility across cloud, on-premises, or hybrid environments, already in production use at companies including ASML and La Banque Postale.