Big 5 AI Vendor Roundup: Week of June 8, 2026

The IPO race and the model race both hit new highs this week, and then the US government did something unprecedented. Anthropic made Claude Fable 5, its first publicly available Mythos-class model, generally available on June 9, only to have Washington order it and the restricted Mythos 5 offline three days later under export-control authority – the first time the federal government has forced a publicly deployed commercial AI model down. OpenAI confirmed its own confidential IPO filing, agreed to acquire coding agent startup Ona, and opened distribution through Oracle. The previously announced agreement between Google and Apple became reality at Apple’s WWDC, where we saw Gemini powering a rebuilt Siri, which will be running on more than a billion iPhones via a deal totaling roughly $1 billion per year. AWS made its Graviton5 chip generally available, borrowed $17.5 billion to fund AI infrastructure, and put Fable 5 on Bedrock at launch. Microsoft had a quiet week marked mainly by a Copilot outage. And in the background, SpaceX went public Friday in the largest IPO in history, setting the benchmark for the Anthropic and OpenAI listings to come.

The two themes that carry over and become clearer are that the frontier briefly went public (Mythos-class capability shipped, then got recalled by the government), and the three trillion-dollar AI listings are moving from filing to pricing.

Anthropic: Fable 5 gets public release – the US government shuts it down

• Claude Fable 5 and Claude Mythos 5 launch (June 9). Anthropic made Claude Fable 5 generally available, its most capable public model to date and the first built on the Mythos-class capability restricted to Project Glasswing partners since April. It scores state-of-the-art across most benchmarks, with the lead widening on longer, more complex tasks. Notably, Anthropic shipped one model as two products: Fable 5 for everyone, with safety classifiers that route flagged cyber, biology, and chemistry requests and attempts to train other models down to Claude Opus 4.8; and Claude Mythos 5, which has the same weights with those safeguards lifted, restricted to vetted cyberdefenders and critical infrastructure operators. Pricing is double that of Opus 4.8 ($10/$50 per million input/output tokens) with a 1M-token context. We covered the model, its pricing math, and the long running agent cost implications in Claude Fable 5 Arrives With Mythos-Class Power and Prices.
• The timing is part of the story. Fable 5 shipped five days after Anthropic's "When AI builds itself" call for a coordinated industry slowdown. Releasing the most capable public model in the world days after warning the technology is moving too fast demonstrates the bind every frontier lab is in where the cautious option and the competitive reality rarely point in the same direction. It also drops as Anthropic moves toward its IPO, where capability leadership will be a headline in the valuation story.
• The US government ordered both models offline (June 12). Three days after launch, Anthropic received an export-control directive from the Commerce Department's Bureau of Industry and Security, signed by Commerce Secretary Howard Lutnick, ordering it to suspend access to Fable 5 and Mythos 5 for any foreign national, inside or outside the US, including Anthropic's own foreign-national employees. Anthropic said compliance required disabling both models for all users worldwide, and it did so the same evening, in what appears to be the first time Washington has forced a publicly deployed commercial AI model offline. The government cited national security and a reported jailbreak but Anthropic pushed back publicly in its statement, suggesting the reference is to a narrow, non-universal jailbreak (asking the model to read a codebase and fix flaws) whose capability it says is widely available from other models including GPT-5.5 and is used daily by defenders. It noted thousands of hours of prelaunch red-teaming with the US government, the UK AI Safety Institute, and external bug bounties produced no universal jailbreak, and warned it would halt all frontier deployments if the same standard applied industry-wide. Access to Anthropic's other models (Opus 4.8 and earlier) is unaffected.
• The impact goes well beyond Anthropic. A single directive took the most capable publicly available model offline for hundreds of millions of users with no notice, and the export control positioning impacts US allies (UK, Canada, Australia, New Zealand) and their nationals. For any enterprise depending on a US frontier model for mission-critical work, that is a continuity risk that may not have been on the risk register a week ago where government action, not just outages or pricing, can remove access overnight. The "build for portability" theme we have been talking about for weeks just added sovereignty to the argument.

OpenAI: IPO filing confirmed, Ona acquisition, Oracle distribution

• Confidential IPO filing confirmed (June 8). OpenAI publicly confirmed it had submitted a confidential draft S-1, one week after Anthropic's filing. CFO Sarah Friar confirmed the submission but said timing is undecided, while reporting points to a possible fall debut above $1 trillion. For enterprise buyers, the takeaway is, as we’ve said in previous roundups, that audited financials and risk disclosures are coming from both leading labs this year.
• Acquiring Ona (June 12). OpenAI agreed to acquire Ona, a coding-agent startup, deepening Codex's position as agentic coding consolidates. It follows the pattern across the industry of labs buying their way into the developer toolchain rather than waiting to build it.
• Oracle distribution (June 11). Enterprises can now access OpenAI models and Codex through existing Oracle cloud commitments, extending the multicloud availability that saw GPT-5.5 and Codex go GA on AWS Bedrock the previous week. OpenAI is pursuing the same multicloud distribution strategy as other AI labs. Separately, the Dreaming background synthesis memory architecture from late May continued rolling out, and we noted when it was announced, persistent automatic profiles warrant a data governance review before broad employee use.

Google: Gemini powers the new Siri on over a billion iPhones

• Gemini-powered Siri at Apple's WWDC (June 8). Apple rebuilt Siri on a custom Gemini model, reportedly around 1.2 trillion parameters, running in Apple's Private Cloud Compute, under a deal worth roughly $1 billion per year. Heavier queries route to Gemini while simple requests stay on Apple's on-device models. It ships in iOS 27 this fall, with the Gemini-class features requiring iPhone 15 Pro or newer. This is a big distribution for Google as Gemini will be behind the default assistant on well over a billion devices. For Google, it offsets the consumer reach gap with ChatGPT, even as it highlights the irony of Google powering a competitor's flagship while building its own.
• The enterprise wrinkle worth watching. iOS 27 Extensions let users set a third-party model as the default assistant, and Claude is among the selectable options. For organizations with iPhone fleets, that means assistant choice, and the associated data handling questions, becomes a manageable policy decision rather than a fixed default.

Amazon: Graviton5 GA, a $17.5 billion AI loan, and Fable 5 on Bedrock

• Graviton5 generally available (June 10). AWS made its fifth-generation Arm processor GA via EC2 M9g and M9gd instances, pitching it for agentic AI. Graviton5 packs 192 cores (the highest CPU core density in EC2), a 5x larger cache, and DDR5-8800 memory, for up to 25% better compute performance than Graviton4. AWS argues that agentic workloads tax CPUs as much as GPUs, since agents coordinate many concurrent tasks, manage context, and shuttle data between accelerators and applications. Meta has committed to tens of millions of Graviton cores for agentic AI, with Uber and Snowflake also deploying. For IT leaders, custom silicon is where AWS competes on the cost side of agentic AI, and the price-performance case is worth modeling for established agent infrastructure.
• Amazon borrows $17.5B for AI build out (June 10). Amazon finalized a $17.5 billion credit facility led by Citigroup, weeks after a large bond sale, to fund continued data center and AI infrastructure expansion. What’s different is that a company that historically funded growth from cash and operating income is now taking on significant debt to keep pace, joining peers in what is shaping up as the most capital-intensive technology build out in decades. This illustrates the point we’ve been making that the cost of staying competitive in AI infrastructure is enormous and rising, and those costs ultimately reflect in the price enterprises pay.
• Claude Fable 5 on Bedrock on day one. AWS made Fable 5 available on Amazon Bedrock and the Claude Platform on AWS at launch. Coming a week after OpenAI's models went GA on Bedrock, it reinforced AWS's position as the neutral, multimodel platform where enterprises run frontier models from both Anthropic and OpenAI under one set of governance and billing controls. The June 12 government directive applies to Fable 5 wherever it runs, including Bedrock, so that access is suspended along with the rest, but the point is that AWS's model-neutrality represents a strategic advantage as the labs compete on applications and the other hyperscalers build proprietary models.

Microsoft: A quiet week and a Copilot outage

• Copilot outage (June 11). After last week's Build 2026 blitz (seven in-house MAI models), Microsoft's notable item this week was a Copilot service disruption that interrupted the productivity layer before being restored. As assistants become the front door to Office work, their uptime is a business continuity concern, not just a feature question. A useful reminder to confirm SLAs and fallback workflows as Copilot embeds deeper into operations.
• Agent 365 licensing note. As of June 1, new Agent 365 purchases require a Microsoft 365 E5 license, a small but notable cost structure change for organizations planning agent rollouts on the Microsoft stack.

Our Take

This was the week the frontier went public and then got recalled. Anthropic put Mythos-class capability into anyone's hands with Fable 5, splitting it into a safeguarded public version and the Project Glasswing trusted partner Mythos 5 version, an architecture that may yet become the template for shipping dangerous capability models. Three days later a government directive took both offline. Set the jailbreak dispute aside, because the precedent is what’s really important: Washington has shown it will use export control authority to pull a deployed commercial model with no notice, reaching not just US users but allied nations and their citizens. For enterprises, AI vendor risk now includes a sovereignty consideration that was theoretical a week ago. The irony is Anthropic spent early June calling for a coordinated pause, and the government effectively imposed a unilateral one it is now fighting.

The themes remain consistent with SpaceX’s largest IPO in history on Friday, valued at over $2 trillion, setting the benchmark OpenAI and Anthropic will be measured against this fall. And Google's Siri deal becomes the biggest distribution channel in consumer technology. The common thread is that the binding question has shifted from "can the model do it" to "can you reliably get access, at what cost, and who can take it away." That is where IT leaders should focus.

What IT leaders should be doing:

• Don't rebuild around Fable 5 yet. The model is offline under government order with no restoration date, so treat any Fable 5 plans as paused. If you ran workloads during the three-day window, keep Opus 4.8 as your production fallback and hold off on committing pipelines to Fable until access and its terms are settled.
• Add sovereignty consideration and uptime risk to continuity planning. Between the Fable takedown and Microsoft's Copilot outage, this week showed access can vanish by government directive or plain disruption. Keep a tested fallback model from a different provider for mission-critical workflows, validate SLAs, maintain manual fallback paths, and favor the neutral platform layers (Bedrock, Foundry) that make switching realistic.
• Treat assistant choice as fleet policy. With Gemini powering Siri and Claude selectable as a default in iOS 27, organizations with iPhones should decide their assistant and data handling policy deliberately rather than accepting defaults.
• Read the S-1s before you sign. With OpenAI, Anthropic, and now publicly traded SpaceX all disclosing, you will soon have audited financials and risk factors on your key vendors. Build that review into fall procurement and renewal timelines.

Want to Know More?

• Claude Fable 5 Arrives With Mythos-Class Power and Prices
• Big 5 AI Vendor Roundup: Week of June 1, 2026
• GPT-5.5-Cyber: The Next Claude Mythos?
• Claude Mythos Preview and Project Glasswing: What IT and Security Leaders Need to Know Now
• Establish Your Adaptive AI Governance Program: From Principles to Practice