Big 5 AI Vendor Roundup: Week of June 1, 2026

The race to public markets and the race to model independence were the big themes of the week. Anthropic confidentially filed for an IPO less than a week after its $65 billion raise, expanded Project Glasswing to roughly 150 new organizations across critical infrastructure, including the EU's cybersecurity agency, and called for a coordinated industry pause if AI systems start improving themselves too quickly. Microsoft used its Build 2026 conference to launch seven in-house MAI models, including its first reasoning model, trained from scratch with no OpenAI distillation, which is clearly its post-OpenAI strategy. OpenAI made its frontier models and Codex generally available on AWS, rolled out a new memory architecture, and was hit with a state lawsuit from Florida. Google moved Gemini 3.5 Flash to general availability while its Search overhaul rattled the online ad world. AWS was the platform beneficiary of OpenAI being generally available (GA) in an otherwise steady week.

The two big themes of the week really started gaining clarity. The first is that the three trillion-dollar AI listings of 2026, Anthropic, OpenAI, and SpaceX, are now in a race to the public markets, and the disclosures will follow. The second is that the hyperscalers are done depending on a single model partner: Microsoft now has its own full model stack, and every major cloud is hosting every major AI lab.

Anthropic: IPO Filing, a Glasswing Expansion, and a Call to Pause

Anthropic followed its record raise with three moves that matter for the long game, including an unusually candid warning about where the technology is headed.

• Confidential IPO filing (June 1). Anthropic confidentially submitted a draft S-1 to the SEC for a proposed IPO less than a week after closing its $65-billion Series H funding at a $965-billion valuation. The confidential route lets the company refine disclosures with regulators before going public, with timing dependent on SEC review and market conditions. Anthropic has not set a share count or price. The filing puts Anthropic ahead of OpenAI in the race to list, with both targeting a possible fall IPO. Together with SpaceX, that makes three potential trillion-dollar AI listings queuing up for the same window. For enterprise buyers, the significance is the same point we identified last week: that a public Anthropic means audited financials, risk disclosures, and a level of transparency we haven’t seen before from an AI vendor. That is information IT leaders have never had when evaluating a frontier lab.
• Project Glasswing expands to ~150 new organizations (June 2). Anthropic extended Claude Mythos Preview access to roughly 150 additional organizations across more than 15 countries, building on the initial group of about 50 partners from April. That first group has reportedly used Mythos to find more than 10,000 high- or critical-severity vulnerabilities. The new group deliberately targets sectors that were underrepresented at launch, including power, water, healthcare, communications, and hardware, many of them software vendors whose code is embedded in critical infrastructure. Anthropic estimates that for most of these organizations a major breach could affect more than 100 million people. Each new partner must meet security requirements before gaining access. Anthropic also admitted the EU's cybersecurity agency, ENISA, into the program, its first EU body, and developed the expansion in collaboration with the US government. Some security researchers see the wider rollout as a step toward an eventual public release of Mythos-class capability.
• The critical infrastructure inclusion is important for IT leaders. Mythos is moving from a small circle of tech and finance firms into the operational technology and infrastructure layer. If your organization runs power, water, healthcare, or communications systems, or depends on vendors who do, the vulnerability discovery capability that has been finding zero-day exploits at scale is now reaching your sector. The patch window compression we’ve talked about in earlier weeks is coming quickly.
• A call for a coordinated pause on frontier development (June 4). In a lengthy piece from the Anthropic Institute titled "When AI Builds Itself," Anthropic argued that frontier labs should build a coordinated, verifiable mechanism to slow or temporarily pause development if AI systems begin improving themselves faster than society can manage the risks. The company put real numbers behind the warning, saying that as of May 2026 more than 80% of the code merged into Anthropic's own codebase was authored by Claude (up from low single digits before Claude Code launched in early 2025), the typical engineer now merges roughly eight times as much code per day as in 2024, and the length of tasks models can complete autonomously has been doubling roughly every four months. Anthropic identifies this as the path toward "recursive self-improvement," where AI can design and train its successors with diminishing human involvement, and says it would support a pause only if other frontier labs, in multiple countries, agreed under verifiable conditions. A unilateral pause, it notes, would just change who leads. The Anthropic Institute plans to convene policymakers, researchers, and other AI labs in the coming months. Many have taken note of the timing of this initiative, considering that calling for an industry-wide slowdown days after filing to go public at a nearly trillion-dollar valuation would also freeze the competitive landscape in favor of today's leaders. Both the safety argument and the incumbent advantage critique can be true.
• What the pause argument means for IT leaders. Set aside the existential messaging and consider what it means for planning. Anthropic is telling the market that capability gains are compounding, not leveling off, and that the pace is being set by how much development AI can now do on its own. For IT leaders, that reinforces a theme from prior weeks: that you should plan for capability and pricing to keep moving quickly, keep evaluations on a short refresh cycle, and treat AI governance as an adaptive, evolving requirement rather than a one-time policy. It also emphasizes the importance of the regulatory alignment discussion, since Anthropic is explicitly inviting policymakers to provide input into how and when development should slow.

Microsoft: Build 2026 and the Post-OpenAI Model Stack Arrives

Microsoft's independence strategy, which we have tracked for weeks as a series of moves, became a product line.

• Seven in-house MAI models, including MAI-Thinking-1 (June 2). At Build 2026, Microsoft's AI superintelligence team, led by Mustafa Suleyman, unveiled seven models developed entirely in-house: MAI-Thinking-1 (its first reasoning model), MAI-Code-1 and MAI-Code-1-Flash, MAI-Image-2.5 (plus a Flash variant), MAI-Transcribe-1.5, and MAI-Voice-2. The headline is that all were trained from scratch on clean, commercially licensed data with no distillation from OpenAI or any third-party model. MAI-Thinking-1 is a midsized model (roughly 35 billion active parameters, ~1 trillion total in a sparse mixture of experts architecture, and a 256,000 token context). Microsoft says it matches Claude Opus 4.6 on coding ability on the SWE Bench Pro benchmark and that raters preferred it to Claude Sonnet 4.6, though these are Microsoft's own benchmark claims. MAI-Code-1-Flash is rolling out to all GitHub Copilot plans this month, MAI-Image-2.5 is live in PowerPoint and Foundry, and the rest are available through Foundry, Fireworks AI, Baseten, and OpenRouter.
• The strategic roadmap becomes clearer. This does not mean Microsoft is breaking with OpenAI. Azure remains OpenAI's primary infrastructure, GitHub Copilot still supports OpenAI models, and M365 Copilot still uses OpenAI capabilities. What changed on June 2 is that Microsoft is now shipping operational alternatives at every tier of the developer stack and standardizing on Foundry and Copilot as the orchestration layer above all model choices. The combination of last month's Inception acquisition talks, the internal Claude Code wind down, and now a complete in-house model family reinforces that Microsoft is building optionality so that it’s not dependent on a single model partner.
• GitHub Copilot moves to usage-based billing (June 1). Copilot's pricing shifted from flat rate request limits to usage-based token billing under a model called AI Credits. Early developer reaction has been almost universally negative. For IT leaders, cost predictability is becoming a real issue as agentic coding tools migrate to consumption pricing and developer costs get much harder to forecast.
• Copilot governance tools and on-device MAI support. Microsoft also introduced new Copilot governance controls and previewed on-device MAI model support for Copilot+ PCs, with a confidential computing fine-tuning API available for public preview in Q3.

OpenAI: Models GA on AWS, New Memory, and a State Lawsuit

It was a mixed week for OpenAI, with a big distribution win and a notable legal development.

• OpenAI frontier models and Codex GA on AWS (June 1). OpenAI's GPT-5.5, GPT-5.4, and Codex are now generally available on Amazon Bedrock, moving out of the limited preview announced earlier this spring. This brings OpenAI to millions of AWS customers through Bedrock's security, governance, and per-token billing, letting enterprises adopt OpenAI through procurement and compliance workflows they already run. It’s a significant distribution expansion and a sign that the major labs increasingly want to be available on every cloud rather than locked to one (covered from the AWS side below).
• Dreaming V3 memory architecture (June 4). OpenAI began rolling out a new ChatGPT memory system that runs background synthesis after conversations, automatically building and updating user profiles without explicit "remember this" commands. OpenAI says it is roughly 5x more compute efficient than the prior system, which is what allows free tier access. It reached Plus and Pro users in the US first. For organizations, persistent automatic memory raises data governance questions worth reviewing before broad employee use.
• Lockdown Mode (June 4). OpenAI introduced an optional high security sandbox that restricts network-heavy tools (live browsing, deep research, file downloads, agent mode) to reduce the risk of data exfiltration by prompt injection and will be available across all tiers via security settings. This is a direct response to the rising prompt injection threat and a useful control for security conscious organizations.
• Florida sues OpenAI (June 1). The Florida attorney general filed an 83-page lawsuit against OpenAI and Sam Altman, the first state-led litigation of its kind, referencing an investigation into ChatGPT's alleged role in a tragic incident. The suit alleges OpenAI prioritized the AI race over safety warnings, while OpenAI has denied responsibility in prior statements. The legal and regulatory cloud over OpenAI remains even with the Musk trial over, which is a risk factor enterprise buyers should keep in an eye on ahead of any IPO.

Google: Gemini 3.5 Flash Goes GA, Search Overhaul Rattles Advertisers

Google had a more operational week as it continued rolling out the I/O lineup.

• Gemini 3.5 Flash generally available. Gemini 3.5 Flash is now GA via the Gemini API and Gemini Enterprise in the Global, US, and EU regions, and enabled by default for Gemini Enterprise users as of June 8. Google positions it as frontier performance for agentic and coding workloads at Flash-series cost and speed. For enterprises building agents, this is now a production option rather than a preview.
• Search overhaul unsettles the ad ecosystem. Following the I/O announcement that Google would reimagine the search box with AI in its biggest change in over 25 years, online advertisers have reacted with alarm at what an AI-mediated search experience does to the traditional page of links and the ad model built on it. This is less an enterprise IT issue than a signal of how disruptive the shift to AI-first interfaces will be for any business that depends on search traffic, and especially for marketing and digital revenue leaders.
• Nano Banana 2 video-to-image. Building on last week's enterprise GA, the Gemini API added video-to-image generation (passing a video file or YouTube URL as context to generate thumbnails, posters, or infographics), supported on the Flash Image model.

Amazon: OpenAI Models Land on Bedrock

AWS' headline this week came from a partner.

• OpenAI GPT-5.5, GPT-5.4, and Codex GA on Amazon Bedrock (June 1). AWS made the OpenAI frontier models and Codex generally available on Bedrock, with built-in security, governance, and pay-per-token pricing. AWS guidance is to use GPT-5.5 for the hardest workloads and GPT-5.4 for best price-performance. With this, Bedrock now hosts frontier models from OpenAI alongside Anthropic's Claude, positioning AWS as the neutral, multimodel platform where enterprises can run any lab's models under one set of controls. That neutrality is increasingly AWS's strategic advantage as the labs themselves compete on applications and the other hyperscalers build proprietary stacks.
• The Bedrock roadmap (AgentCore payments, Agent Toolkit, and now OpenAI models GA alongside the Claude Platform) is what to watch for IT leaders standardizing agent infrastructure.

Our Take

This was the week the AI IPO race heated up. Anthropic's confidential filing, on the heels of its $65 billion raise and ahead of OpenAI, sets up a fall window where the market finally gets audited financials from a frontier lab. For IT leaders, it’s the most useful development of the cycle, moving from valuation headlines and run-rate claims to actual disclosure. The flip side is that the legal and regulatory implications become front and center, with Florida's lawsuit against OpenAI a reminder that these companies carry risk factors that will be spelled out in their prospectuses.

The independence is the other big theme. Microsoft shipping seven in-house models, OpenAI going GA on AWS, and every major cloud now hosting every major lab all point to the end of the era of exclusive model vendor lock-in. The strategic question for enterprises is no longer “which vendor do I bet on” but “how do I stay portable across them.”

Sitting underneath both races is Anthropic's pause argument, which is the most candid statement yet from a frontier lab that the pace is accelerating, not stabilizing. Whether you read it as genuine caution, savvy pre-IPO positioning, or both, the operational takeaway is that the vendors themselves are telling you capability and pricing will keep moving fast. That is an argument for shorter evaluation cycles and adaptive governance, not annual reviews.

What IT leaders should be doing:

• Use the coming IPO disclosures. When Anthropic and OpenAI file public S-1s, read the risk factors and financials before signing or renewing multiyear commitments. You will finally have audited data on revenue concentration, compute costs, and legal exposure. Build a review of those disclosures into your procurement timeline this fall.
• Assess MAI models if you are a Microsoft shop. With MAI-Code-1-Flash rolling into all GitHub Copilot plans and MAI-Image-2.5 in PowerPoint and Foundry, Microsoft's in-house models are now in your environment whether you sought them out or not. Benchmark them against Claude and OpenAI on your own workloads rather than relying on vendor benchmark claims and confirm which model is actually serving each Copilot instance.
• Reexamine coding agent costs under consumption pricing. GitHub Copilot's shift to token-based AI Credits is the second consumption pricing move in two weeks. Model your per-developer cost exposure under usage-based billing and set guardrails before rolling out agentic coding tools broadly.
• Treat multicloud, multimodel as the default architecture. With OpenAI now GA on AWS alongside Claude, and Microsoft, Google, and AWS all hosting multiple labs, portability is achievable. Avoid architectures that commit you to one model provider; use the neutral platform layers (Bedrock, Foundry) that let you switch.
• Review data governance for automatic memory. OpenAI's Dreaming V3 builds persistent user profiles without explicit prompts. Before enabling it for employees, confirm how that data is stored, who can access it, and whether it meets your data residency and privacy requirements.
• If you run critical infrastructure, engage on Mythos now. With Glasswing expanding into power, water, healthcare, and communications and ENISA now in the program, the vulnerability discovery capability is reaching operational technology sectors. Ask whether your organization or your key vendors have access and plan for shorter patch windows as AI-discovered vulnerabilities accelerate.

Want to Know More?

• Big 5 AI Vendor Roundup: Week of May 25, 2026
• Big 5 AI Vendor Roundup: Week of May 18, 2026
• GPT-5.5-Cyber: The Next Claude Mythos?
• Claude Mythos Preview and Project Glasswing: What IT and Security Leaders Need to Know Now
• Anthropic's "The Briefing: Financial Services" Event Was Different in the Best Way
• Establish Your Adaptive AI Governance Program: From Principles to Practice