Big 5 AI Vendor Roundup: Week of June 15, 2026

The fallout from the US government's shutdown of Anthropic's Fable 5 and Mythos 5 dominated the week and followed the AI labs CEOs all the way to the G7. At a working AI lunch at the summit in Évian-les-Bains, France, on June 17, the heads of Anthropic, Google DeepMind, and OpenAI sat with President Trump and G7 leaders while the elephant in the room was a model Washington had switched off five days earlier. Both models remain offline worldwide a week after the order, with no confirmed restoration date. Around that, Google's Gemini 3.5 Pro is still in preview as its "next month" promise runs down, OpenAI teased GPT-5.6 and notched a second courtroom win over Elon Musk, AWS shipped a FinOps agent and put Google's Gemma 4 on Bedrock, and Microsoft's Satya Nadella warned that AI concentration could "hollow out entire industries."

One of the themes we have tracked for weeks hardened into the week's defining issue: the binding question for buyers is no longer whether a model can do the work, but whether you can reliably get access to it, at what cost, and who can take it away. This week, the answer to "who can take it away" turned out to include the US government.

Anthropic: Fable 5 and Mythos 5 stay dark as the export ban reaches the G7

• Day seven and still offline. The Commerce Department's June 12 export control directive forced Anthropic to disable Fable 5 and Mythos 5 for all users worldwide, and as of June 19 neither has returned. Anthropic's other models (Opus 4.8, Sonnet, Haiku) are unaffected and remain the working fallback. The company maintains the trigger was a narrow, non-universal jailbreak whose capability exists in other public models. The directive reportedly followed Amazon CEO Andy Jassy alerting Treasury Secretary Scott Bessent and other senior officials that Amazon researchers had used Fable 5 to extract cyberattack-useful information, which is an awkward turn given Amazon is one of Anthropic's largest investors. When asked about the reporting, Amazon didn’t deny it. Anthropic has sent engineers to Washington to negotiate, and its international managing director said at a June 18 event he was "very confident" the models would return "in the coming days." The subscriber refund deadline for access purchased June 9-14 is June 20.
• The G7 turned into a referendum on the "kill switch." At the June 17 working lunch, the export ban overshadowed the agenda. Dario Amodei and Google DeepMind's Demis Hassabis pressed for a US-led AI coalition, while leaders discussed a "trusted partners" approach to restore allied access to restricted US models. The reactions were candid, as French President Macron warned that nobody will buy US AI if they fear it can be switched off at any moment, and Canadian Prime Minister Mark Carney framed it as a lesson in overreliance, saying allies "will have done something wrong if we just accept this, don't take the lesson, don't build out and diversify." President Trump, from the summit sidelines on June 18, said negotiations with Anthropic were "going fine," which was his first public comment on the ban.
• Anthropic opened a Seoul office (June 18). The launch, intended to mark Korean expansion, was instead dominated by export ban questions, underscoring how directly the sovereignty issue now follows the company internationally. Anthropic reiterated its revenue has grown to roughly $47 billion run-rate from $9 billion at the end of 2025.
• A telling reversal. Anthropic pulled back a planned Claude Agent SDK billing overhaul on June 15, the day it was set to take effect. This is a good reminder that even amid the larger crisis, the pricing model for agentic tooling remains in flux.

Google: Gemini 3.5 Pro still in preview as the "next month" clock runs down

• The flagship is late. Gemini 3.5 Pro, unveiled at I/O on May 19 with a June general availability target and Sundar Pichai's "give us until next month" promise, remains in limited Vertex AI preview for select enterprise customers as of June 19. The targeted feature set is significant, including a 2-million-token context window and a "Deep Think" reasoning mode, absorbing the use cases Google previously routed to its Ultra tier, but with roughly ten days left in the month, the window is closing. Gemini 3.5 Flash remains the public 3.5 model. For IT leaders, the practical guidance is to treat the announced specs as a target until general availability and not to build evaluations around a model that is not yet broadly testable.
• At the G7, Hassabis pushed for coordination. Alongside Amodei, Hassabis advocated a US-led coalition approach to frontier AI, positioning Google DeepMind within the governance conversation rather than just the product race.
• Gemma 4 reached Amazon Bedrock. Google DeepMind's open weight Gemma 4 family became available on a competitor's platform this week (covered in the AWS section), extending Google's model reach beyond its own cloud.

OpenAI: GPT-5.6 teased, a second legal win over Musk, and a testing forum pitch at the G7

• GPT-5.6 on the horizon. OpenAI's chief scientist described GPT-5.6 as a meaningful improvement over GPT-5.5, with a late-June launch widely expected but no official date, specs, or system card yet. As with Google's Pro delay, the takeaway for buyers is to wait for general availability and a published system card before committing.
• Federal judge dismisses xAI's trade secret suit against OpenAI. A judge threw out Elon Musk's xAI trade secret case with prejudice, ruling that recruiting conversations are not trade secrets. This is OpenAI's second legal win over Musk in roughly five weeks following the May jury verdict. The cumulative effect is that OpenAI's most prominent legal complications are evaporating as it moves toward its IPO.
• At the G7, Altman proposed a global testing forum. Where Amodei and Hassabis pushed a US-led coalition, Sam Altman called for a global testing forum with common safety benchmarks. This is a more multilateral framing, and one that looks much different in a week when unilateral US action took a frontier model offline.

Amazon: A FinOps agent, Gemma 4 on Bedrock, and the Fable 5 revocation

• AWS FinOps Agent (preview). AWS introduced an agent aimed at cloud cost management in its June 15 roundup. This is great timing as agentic workloads drive unpredictable consumption costs, which is a concern we have raised repeatedly. A managed tool to surface and control AI and cloud spend is a timely response to the budgeting problem IT leaders and CFOs are increasingly wrestling with.
• Gemma 4 on Bedrock. Google DeepMind's Gemma 4 open models arrived on Bedrock in three variants, including a dense 31 billion parameter model with a 256K token context window. Hosting Google's open models alongside Anthropic's and OpenAI's reinforces AWS's multi-model neutrality, which is increasingly its strategic position as the AI labs compete on model capability.
• Kiro Pro Max. AWS expanded its Kiro coding agent tiers, continuing to consolidate its developer tooling roadmap around Kiro as Q Developer winds down.
• The Fable 5 footnote. AWS confirmed that on June 12 Anthropic asked it to revoke Fable 5 and Mythos 5 access on Bedrock to comply with the export directive. Recall that access to those models required opting into Anthropic's 30-day data retention policy via a Data Retention API, which is a governance consideration for any regulated workload regardless of the ban.

Microsoft: Nadella warns AI concentration could "hollow out entire industries"

• A pointed warning from the top (June 14). In a lengthy post, Satya Nadella argued that the industry may be building toward an outcome "the political economy will simply not tolerate," comparing unchecked AI concentration to the way early globalization hollowed out industrial economies through outsourcing. His point is that backlash to AI concentration could invite regulation that reshapes the economics for everyone, foundation model AI labs in particular. Coming the same week a government export order took a leading model offline, the warning is timely, and it illustrates Microsoft's own interest in not being dependent on any single model provider. Otherwise, Microsoft had a quiet week following its Build 2026 push.

Our Take

This was the week AI sovereignty stopped being a talking point and became a procurement consideration. A single US directive kept the most capable public model in the world offline for a full week worldwide, and the issue brought three AI CEOs onto the stage at the G7, where allied leaders openly questioned whether they can depend on American models that can be switched off without warning. Set aside the merits of the jailbreak dispute, which remains unresolved, the real lesson is that government action now matches outages, pricing, and capacity as a reason your access to a frontier model can disappear. Macron's warning that nobody will buy AI they fear can be unplugged and Carney's call to "build out and diversify" are the customer voice on this, and enterprise buyers should hear it the same way.

Everything else this week pointed in a consistent direction. Google's and OpenAI's flagships are running late, which argues against building plans around unreleased models. AWS shipped a cost control agent precisely because agentic spend is hard to predict. And Nadella warned that the concentration underpinning the whole race may provoke a regulatory response. The common thread is concentration and control, and it boils down to who holds the capability, who can revoke it, and what it costs to stay portable.

What IT leaders should be doing:

• Treat the Fable 5 ban as a live continuity drill. If you ran anything on Fable 5 or Mythos 5, you have now experienced an overnight loss of access. Keep Opus 4.8 or another provider's model as a tested production fallback, and do not commit pipelines to Fable until access and its terms are settled.
• Make portability and sovereignty explicit in contracts. Ask vendors what happens to your access under an export directive or a government order, where your data and inference are located, and how quickly you could fail over to an alternative. Favor neutral platform layers (Bedrock, Foundry, Vertex) that make switching quickly realistic.
• Don't plan around unreleased models. With Gemini 3.5 Pro still in preview and GPT-5.6 not yet shipped, evaluate only what is generally available and has a published system card. Treat announced specs as targets.
• Get ahead of agentic cost control. Tools like the AWS FinOps Agent exist because consumption-based agent spend is nearly impossible to forecast. Put cost guardrails and monitoring in place before scaling agent deployments, not after the first bill shock.
• Watch the data retention fine print. Mythos-class access on Bedrock required opting into 30-day data retention. Confirm the data handling terms of any frontier model before putting regulated or sensitive workloads through it.

Want to Know More?

• Big 5 AI Vendor Roundup: Week of June 8, 2026
• Claude Fable 5 Arrives With Mythos-Class Power and Prices
• Claude Mythos Preview and Project Glasswing: What IT and Security Leaders Need to Know Now
• GPT-5.5-Cyber: The Next Claude Mythos?
• Establish Your Adaptive AI Governance Program: From Principles to Practice