It is Monday, June 16, 2026. The US government shut down Anthropic’s two most powerful models four days ago and they are still offline. A Chinese lab dropped a one-trillion-parameter open-source coding model the same day. And Google’s flagship model remains in the waiting room while the clock on a public promise ticks toward zero. Welcome to this week.
The US government pulled Anthropic’s two best models and they are still gone
This is the story the AI industry cannot stop talking about and it is still unresolved as of this morning. On the evening of June 12, 2026 — three days after Anthropic launched Claude Fable 5 to the public — the US Department of Commerce sent a letter to CEO Dario Amodei ordering the immediate suspension of access to Claude Fable 5 and Claude Mythos 5 by any foreign national, inside or outside the United States. Because Anthropic has no way to filter foreign nationals from domestic users in real time, it shut both models down for everyone worldwide. Claude Opus 4.8, Sonnet 4.6, and Haiku 4.5 are unaffected and remain online.
The stated reason: the government believes it found a method to jailbreak Fable 5. Anthropic published its own account of what the jailbreak actually is. According to the company, the method consists of asking the model to read a specific codebase and identify software vulnerabilities — a capability Anthropic says is already widely available in GPT-5.5 and is used daily by the security professionals who defend systems. Anthropic called the directive a misunderstanding and said applying the same standard industry-wide would amount to a halt on all new frontier model launches.
The political backstory matters here. In March 2026, the Defense Department classified Anthropic as a supply-chain risk after the company refused to remove safety guardrails — specifically, a prohibition on building fully autonomous weapons systems. Anthropic is challenging that designation in court. David Sacks, co-chair of the President’s Council of Advisers on Science and Technology, revealed on June 13 that the administration gave Anthropic a choice before the ban: fix the jailbreak or pull the model. Anthropic disagreed that the jailbreak warranted a recall. The government disagreed with their disagreement. Fable 5 has now been offline for four days.
Prediction markets give roughly 40 percent odds that Fable 5 is restored for US users before June 20. Manifold and Polymarket both reflect trader uncertainty about whether the resolution involves a technical patch, a policy carve-out, a legal challenge, or a prolonged standoff. This is the first time in history a US government directive has forced a publicly deployed frontier AI model offline.
Read the coverage: Anthropic’s official statement on the suspension | Time on the ban and what it reveals about the government’s AI strategy | MarkTechPost on the full timeline and community reaction | Snyk on what the jailbreak actually was and why security teams should care | ExplainX on the realistic paths to restoration and what developers should do now
China dropped a one-trillion-parameter open-source coding model the same day Fable 5 went dark
The timing is not subtle. On June 12 — the same evening Anthropic’s most capable models went offline — Beijing-based Moonshot AI released Kimi K2.7 Code, a one-trillion-parameter mixture-of-experts coding model available as open weights on Hugging Face under a Modified MIT license. API pricing is $0.95 per million input tokens and $4.00 per million output tokens. Claude Opus 4.8 charges $5.00 input and $25.00 output. GPT-5.5 charges $1.50 input and $30.00 output. The cost gap is not subtle either.
K2.7 Code is built on the Kimi K2.6 architecture with a coding-specific training run layered on top. The claimed benchmark improvements over K2.6: plus 21.8 percent on Moonshot’s own Kimi Code Bench, plus 11.0 percent on Program Bench, plus 31.5 percent on MLS Bench Lite, and approximately 30 percent lower reasoning token consumption per task. The important caveat is that every one of those numbers comes from Moonshot’s internal benchmark suite rather than independent leaderboards like SWE-bench Verified. Third-party validation has not yet landed.
The model has a 256,000-token context window, forces reasoning mode on by default, and ships with a MoonViT vision encoder for multimodal inputs. Moonshot has maintained a two-to-three month release cadence since launching the original K2 in July 2025, and has publicly signaled that K3 — reportedly three to four trillion parameters — is the next major architecture jump. For the developers who were mid-deployment on Fable 5 when it went dark Friday night, K2.7 Code arrived at a convenient moment.
Read the coverage: TechCrunch on the broader open-source coding model race | Build Fast With AI’s detailed K2.7 Code benchmark review | Flowtivity’s hands-on review comparing K2.7 against GPT-5.5 and Claude Opus 4.8 | CryptoBriefing on Moonshot’s full K2 release history and what K3 might look like
Google’s flagship model is now 28 days past its promise date and June is running out
At Google I/O on May 19, Sundar Pichai told the audience that Gemini 3.5 Pro would ship the following month. The audience reportedly groaned. It is now June 16. The model has not shipped. Fourteen days remain in June.
What is confirmed about the model: a two-million-token context window, a Deep Think reasoning mode, and frontier multimodal capability across text and images. It is designed to handle the workloads Google previously routed to Gemini Ultra — complex long-horizon reasoning, deep agentic tasks, and the hardest coding benchmarks. Pricing is expected in the range of $2 to $4 per million input tokens and $12 to $25 per million output tokens, placing it in direct competition with Claude Opus 4.8 and GPT-5.5. The model has been in internal use and limited enterprise preview on Vertex AI since before I/O.
The competitive pressure has increased since the original promise. Apple announced last week that the new Siri runs on a $1 billion-per-year Gemini license. That deal covers two billion active iPhone users starting this fall, and the model powering it — Gemini 3.5 Pro in all but name — has not finished shipping. Prediction markets at Polymarket are pricing a late June release, with a meaningful probability on the model missing the month entirely. If it does, Google will have made a public commitment to developers at its flagship annual conference and missed it by more than six weeks.
Read the coverage: TechTimes on the confirmed specs and what Deep Think reasoning delivers | Polymarket current odds on the release window | The most realistic breakdown of dates, pricing, and how it stacks up against the current frontier
SPCX closed its first week up 25 percent and the Nasdaq-100 door is already opening
SpaceX ended its first full week as a public company at approximately $169 per share, a 25 percent gain over its $135 IPO price. The MSCI global indices began their inclusion review process over the weekend. Nasdaq-100 membership — which would trigger automatic buying from every index fund and ETF that tracks it — is estimated to be roughly ten trading days away from the first eligibility window.
The number that is drawing the most scrutiny from analysts is the xAI division’s cash burn. SpaceX’s S-1 disclosed that xAI consumed $14 billion in cash against $3.2 billion in revenue in 2025, a net drain of more than $10 billion on the AI side of the business. The profitable core remains Starlink broadband at $11.4 billion revenue and $4.4 billion operating income in 2025. Every dollar of AI ambition at xAI is currently being funded by satellite internet subscriptions from nine million global customers. Jensen Huang told clients this week that buying into the SpaceX IPO was like getting early access to Amazon, Google, or Meta at their listings. That is an optimistic comparison. It is also the comparison that is moving institutional money.
Read the coverage: NPR on the debut and market reception | CNBC’s full first-day live coverage and analyst commentary
Also worth reading today
- Anthropic published its most significant AI safety paper of the year on June 4, proposing a coordinated global AI pause mechanism for frontier models if capability thresholds are crossed. The paper has drawn both praise from safety researchers and skepticism from the AI developer community about enforceability. (Build Fast With AI June 15 roundup)
- Daniela Amodei explained publicly why compute costs are forcing Anthropic toward an IPO. The company is paying SpaceX $1.25 billion per month through May 2029 for compute access — $15 billion per year to a single vendor. That number appears in Anthropic’s confidential S-1 and defines the company’s margin problem heading into its public offering. (Build Fast With AI June 15 roundup)
- The EU AI Act’s first major enforcement deadline is now 14 days away. Any company with a frontier model deployed in Europe needs transparency reports, governance documentation, and systemic risk assessments in place by June 30. The clock is running and the list of companies that are not ready is longer than regulators are publicly acknowledging. (Build Fast With AI June 15 roundup)
That is your Monday. Anthropic’s two best models are still dark, a Chinese lab filled the vacuum with open-source weights the same night, Google owes the world a model it promised four weeks ago, and SpaceX is up 25 percent in its first week. The circus does not take weekends off. See you tomorrow.