Big 5 AI Vendor Roundup: Week of June 22, 2026

Two weeks after the US government pulled Anthropic's Fable 5 and Mythos 5 offline, the US government gave Anthropic permission to rerelease Mythos 5 to about 100 US government agencies and private companies that operate and defend critical infrastructure. Fable 5 remains offline and we also found out this week why Washington reacted so hard. An unnamed US official confirmed to the Associated Press that Anthropic’s Mythos model found vulnerabilities in classified government systems during a joint testing exercise. While Anthropic fought that fire, OpenAI spent the week leaning into the exact capability that got Anthropic in trouble, shipping its GPT-5.5-Cyber model to vetted defenders and unveiling a custom inference chip with Broadcom. Google lost two of its most prominent researchers – one to Anthropic and one to OpenAI – and watched its stock fall about 6%. AWS used its New York Summit to ship a stack of agent infrastructure and added xAI’s Grok to Bedrock. Microsoft kept quietly stitching Azure and Copilot into a single governed control plane.

If last week was about AI sovereignty, this week was about its consequences. Frontier cyber capability is now a government matter, the best researchers are consolidating at Anthropic and OpenAI, and Google spent the week explaining departures instead of shipping its delayed flagship.

Anthropic: Fable 5 still offline, Mythos 5 rereleased to “trusted partners,” and we now know why Washington reacted

• Two weeks later, Mythos 5 is rereleased. Fable 5 and Mythos 5 went offline on June 12 under the Commerce Department’s export control order, and on June 26 the US Commerce Secretary sent a note to Anthropic allowing it to release Mythos 5 to 100 organizations that the government approved. Fable 5 remains offline with no end to the export restriction in sight. The free access window Anthropic had promised through June 22 expired without a single hour of real availability, and the models are now nominally on usage-based pricing that nobody can actually use. Anthropic’s other models (Opus 4.8, Sonnet, Haiku) are unaffected and remain the practical fallbacks. Executives have said for over a week that access would return “in the coming days,” but there is still no restoration date for Fable 5, and prediction markets are pricing better-than-even odds that it slips past July 1.
• The NSA testing revelation. An unnamed US official told the Associated Press that Anthropic’s Mythos model identified vulnerabilities in highly sensitive, classified US government systems during a testing exercise run with US intelligence agencies under Project Glasswing. Senator Mark Warner had referenced the same exercise in a congressional hearing, relaying that NSA chief Joshua Rudd told him Mythos “broke into almost all of our classified systems, not in weeks, but in hours.” The caveat, from the AP’s own source, is that Mythos found the vulnerabilities quickly but that does not mean it could exploit them in that time, so Warner's “broke into” line is stronger than the reporting supports. The New York Times separately reported that the NSA lost access to Mythos once the ban took effect, which means the government’s own defenders are now without a tool they had been using. This is the clearest explanation yet for why the administration moved so aggressively. The worry was never a narrow jailbreak. It was the model’s raw offensive cyber ability.
• Anthropic keeps winning the talent war. John Jumper, who shared the 2024 Nobel Prize in chemistry for AlphaFold, announced he is leaving Google DeepMind for Anthropic after nearly nine years. He joins ahead of an Anthropic science event on June 30, and his background in protein structure and scientific computing hints at where Anthropic wants to go beyond coding and security. Landing a Nobel laureate in the same week it is fighting the US government says something about the company’s pull right now.

OpenAI: Leaning into cyber while building its own chip

• GPT-5.5-Cyber goes live, plus an open-source push. OpenAI launched the full version of GPT-5.5-Cyber on June 22 through its Daybreak program, gated to vetted organizations under the Trusted Access for Cyber program (Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, Zscaler, and government partners). OpenAI says it recorded the highest CyberGym score of any single model to date. It paired that with Patch the Planet, a Daybreak initiative with security firm Trail of Bits to help overstretched open-source maintainers triage AI-generated vulnerability findings before they pile up. The contrast with Anthropic is hard to miss. In the same week the government took Anthropic's cyber model offline, OpenAI shipped its own to defenders and noted it has been in ongoing dialogue with Washington about its cyber approach and upcoming releases. Two labs, two very different outcomes on frontier cyber.
• A custom inference chip with Broadcom. On June 25, OpenAI and Broadcom unveiled Jalapeño, a custom accelerator built specifically for LLM inference, with first deployment targeted for late 2026. That puts OpenAI alongside Google (TPU), Amazon (Trainium), and Microsoft (Maia) in designing its own silicon to control the cost and supply of inference, which is the real constraint as these products scale. Broadcom CEO Hock Tan called it the start of a multigeneration roadmap.
• A tuck-in acquisition and a marquee hire. OpenAI acquired Astral, the maker of the widely used Python tools uv and Ruff and is folding them into the Codex stack. Separately, Noam Shazeer, co-author of the original Transformer paper and a Gemini co-lead, left Google to join OpenAI as lead for architecture research. Sam Altman called him a hire he had wanted since OpenAI’s earliest days.

Google: Losing researchers and explaining itself

• Two big departures and a market reaction. Within a few days Google DeepMind lost Shazeer to OpenAI and Jumper to Anthropic, and Alphabet shares fell about 6% on June 22 as the back-to-back exits raised real questions about whether DeepMind can stay in the lead pack. Google’s top models, Gemini 3.5 Flash and 3.1 Pro, have been sliding down the benchmark leaderboards, and the departures fed a growing sense that the company is playing not to lose rather than playing to win.
• Gemini 3.5 Pro is still not generally available. The flagship Google promised “next month” at I/O is still in limited Vertex preview with June nearly gone. The guidance from last week stands: Evaluate what is actually shipping, not what has been announced.

Amazon: A summit full of agent plumbing, and Grok joins Bedrock

• New York Summit launches. At its NY Summit, AWS rolled out a set of agent-focused capabilities: fully managed knowledge bases in Bedrock for enterprise retrieval (native connectors, smart parsing, an agentic retriever); a managed web search tool for AgentCore that grounds agent answers in cited web results with zero data egress from your environment; and Lambda MicroVMs, a serverless primitive that gives agents isolated sandboxes with state preserved up to eight hours. The zero-egress web search is the standout for regulated buyers, since it adds live grounding without sending data outside your AWS account.
• Grok 4.3 on Bedrock, and price cuts. AWS added xAI’s Grok 4.3 to Bedrock and announced a round of price reductions. Bedrock now hosts Anthropic, OpenAI, Google’s Gemma, and xAI under one set of controls, which keeps reinforcing AWS’s position as the place to run any lab’s models without betting your stack on one of them.

Microsoft: Assembling the control plane

• Azure and Copilot as one governed layer. Microsoft’s quieter week was mostly strategic positioning. Executives and analysts described a push to turn Azure, Copilot, GitHub, Microsoft 365, and security into a single AI control plane, where you define a model policy once (which models can touch which data), and it applies across every Copilot surface, with a federated catalog that routes each request to the cheapest or most private endpoint. It is the governance and cost story IT leaders keep asking for, though much of it is still in preview.
• SMB Copilot pricing firms up. Microsoft is turning its Business Standard and Premium “with Copilot” promos into permanent SKUs on July 1, at $23.50 and $32 per user per month, with a standalone Copilot Business option around $18. It’s worth modeling now if you are planning Copilot across a smaller organization.

Our Take

The real story this week is that frontier cyber capability has become a government-controlled asset. The AP’s confirmation that Mythos found holes in classified systems, and the NYT’s report that the NSA lost access to the model once the ban hit, show this was never really about a narrow jailbreak. It is about a model that can find serious vulnerabilities faster than human teams and about a government that has decided it cannot let that capability move freely. Even with the limited Mythos 5 rerelease, the US government is defining who can use it, and it is only American organizations. OpenAI is working the same terrain more smoothly, shipping GPT-5.5-Cyber to vetted defenders while staying in close contact with Washington. The takeaway for buyers is that access to the most capable security tooling is going to come with strings attached, from vendors or governments or both, and you should assume your access to it can change when policy does.

The quieter story is talent, and it is moving in one direction. A Nobel laureate and the co-author of the Transformer paper both left Google in the same week, for Anthropic and OpenAI. One week of hires does not decide a race, but the pattern over the past two months is real, and it lines up with where benchmark leadership and enterprise momentum already sit. Google has the capital and distribution to recover, but this week it was explaining exits rather than shipping the flagship it promised for June.

What IT leaders should be doing:

• Plan as if Fable 5 is not coming back soon. Two weeks in with no date, keep Opus 4.8 or another provider’s model as your tested production fallback, and do not rebuild pipelines around Fable until access and pricing are actually settled.
• Treat advanced security AI as access-controlled by default. GPT-5.5-Cyber is gated to vetted defenders, and Mythos is restricted and now banned. If you want this class of capability, start the vetting and partner conversations early and decide what you will use in the meantime.
• Expect every frontier vendor to meter inference like a utility. With OpenAI now designing its own chip alongside Google, Amazon, and Microsoft, the whole industry is squeezing inference cost, and that will keep showing up as consumption-based pricing. Put cost monitoring in place before you scale, not after.
• Use platform neutrality to stay portable. With Bedrock now hosting Anthropic, OpenAI, Google, and xAI behind one set of controls, lean on neutral platform layers so you can move between models as capability, price, and availability change.
• Do not pick a lab based on last week's leaderboard. Talent and benchmark leadership are shifting month to month right now. Run your evaluations on a regular cadence instead of locking in based on who looks ahead today.

Want to Know More?

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