Week in AI: June 8–14, 2026 (Week 24)

This week’s news: June 8–14, 2026 (Week 24)

The week opened with Apple’s WWDC conference, where Apple unveiled a new Siri built on Google Gemini models, and continued with the launch of Anthropic’s most powerful publicly available model, Claude Fable 5. Just three days later, Anthropic had to shut the model down worldwide on the orders of the US government, the first time the US government has forced a developer to pull its flagship model from the market. The financial world was dominated by SpaceX’s record-breaking stock market debut, a company that now also includes xAI and its Grok chatbot, while OpenAI confirmed that it too is preparing to go public. Meanwhile, OpenAI is buying the German startup Ona for its coding agents, and France’s Mistral is in talks over another multibillion investment.

New Models

Anthropic launched Claude Fable 5 and Claude Mythos 5

On June 9, Anthropic released two models from its new top tier, Mythos, which sits above the existing Opus line. Claude Fable 5 is intended for everyday users and businesses, while Claude Mythos 5 is the same model with relaxed safety guardrails and was made available only to selected partners (Anthropic).

The difference between the two versions lies in so-called classifiers, auxiliary models that watch for sensitive queries. With Fable 5, requests related to cybersecurity, biology and chemistry are automatically redirected to the weaker Claude Opus 4.8; according to the company, this affects fewer than 5% of sessions.

Price: $10 per million input and $50 per million output tokens (a token is roughly a piece of a word), which Anthropic says is less than half the price of the previous version, Mythos Preview.
Availability: through the Claude API and in the Pro, Max, Team and Enterprise plans; on launch day the model also appeared in Amazon Bedrock (Info-Tech).
Mythos 5 went to partners in the Project Glasswing program, that is, companies and institutions that protect critical infrastructure and defend against cyberattacks.
Anthropic reports top results in programming, data analysis and image work; among other things, the model finished the game Pokemon FireRed based solely on what it “saw” on the screen.

Google: Gemini 3.5 Live Translate translates speech in near real time

On June 9, Google introduced the audio model Gemini 3.5 Live Translate, which translates spoken language into spoken language in more than 70 languages. The translation lags only a few seconds behind the speaker and preserves their intonation, pace and pitch (Google).

Developers can access the model in public preview in the Gemini Live API and Google AI Studio.
Businesses will start trying it in Google Meet, and regular users in the Google Translate app for Android and iOS.
The generated audio carries an invisible SynthID watermark that reveals it is a synthetic voice.

Moonshot AI released the open model Kimi K2.7 Code

On June 12, the Chinese company Moonshot AI published Kimi K2.7 Code, a coding-focused model whose weights (the trained model itself) anyone can download from Hugging Face. It uses a MoE (mixture of experts) architecture with a trillion parameters, of which only 32 billion are active for each query (Digital Applied).

According to the maker, the model uses about 30% fewer tokens when “thinking” than the previous Kimi K2.6. However, the reported results come only from the company’s own tests.

Products and Agents

Apple at WWDC: Siri AI powered by Gemini models

At the WWDC keynote on June 8, Apple showed off its redesigned assistant, Siri AI. The company announced its partnership with Google back in January; now it showed what it will deliver in practice: Siri understands context across apps and works with what is on the screen as well as what the camera sees (Engadget).

Siri will also get its own chatbot-style app that answers questions, writes text, generates images and analyzes files.
At first it will be English-only and will not be available in iOS 27 and iPadOS 27 in the European Union; Apple cites the rules of the Digital Markets Act (DMA). It will work on Mac and Vision Pro in the EU.
New Apple Intelligence features are coming to Safari, Mail, Messages, Calendar and Photos; all images created or edited by artificial intelligence will carry a SynthID watermark.
iOS 27 and macOS Golden Gate will ship in the fall. For Tim Cook, this was his last WWDC as head of the company; he is set to be succeeded by John Ternus on September 1.

Gemini outage

On June 10, Gemini suffered an outage lasting several hours. Users reported errors 1076 and 1099, and Gemini features in Google Workspace were affected as well. Gemini app lead Josh Woodward confirmed the outage, and according to Google’s status page, the problem was resolved around 10:30 Pacific time (9to5Google).

Other news

June 10: Oracle Cloud customers will be able to use their prepaid credits on OpenAI models and the Codex coding agent without signing a new contract (OpenAI).
June 13: Databricks open-sourced Omnigent under the Apache 2.0 license. It is a “meta-harness”, a layer that connects and orchestrates various agents such as Claude Code or Codex, keeps track of their budgets and permissions, and lets users share sessions with colleagues (Databricks).

Business and Investment

OpenAI filed a confidential IPO application

On June 8, OpenAI announced that it had confidentially submitted a draft S-1, the application to go public, to the US regulator. It has not yet set a date: “it may take a while,” the company said. Rival Anthropic filed a similar application a week earlier, on June 1 (NBC News).

For a sense of scale: in March, after a $122 billion investment, OpenAI was valued at $852 billion, while Anthropic was valued at $952 billion in its latest round.

SpaceX and xAI go public: the largest IPO in history

SpaceX, which since February has also included xAI and its Grok chatbot, priced its shares at $135 on June 11 and raised $75 billion by selling 555.6 million shares. It was the largest initial public offering (IPO) in history (SiliconANGLE).

The shares trade on the Nasdaq under the ticker SPCX; on the first trading day, June 12, they rose about 19%, and by the end of the day the company was valued at roughly $2.1 trillion.
According to the prospectus, the space and AI side of the company generated $1.4 billion in revenue in the first quarter but an operating loss of $3.1 billion. Only the Starlink satellite internet business is profitable.
Elon Musk retains 82.4% of the voting rights.

OpenAI is buying Ona (formerly Gitpod)

On June 12, media reported that OpenAI is buying the German startup Ona, formerly known as Gitpod, founded in 2020 in Kiel. Ona runs secure cloud development environments in which AI agents can work (The Decoder).

Thanks to the acquisition, Codex should be able to handle tasks lasting hours or even days, even when the user’s laptop is closed, and do so directly in the customer’s cloud. According to OpenAI, more than 5 million people use Codex every week; the price of the deal was not disclosed, and the transaction still needs regulatory approval.

More business news

June 10: Amazon secured a $17.5 billion credit facility led by Citigroup to fund the construction of data centers and AI infrastructure (TechBuzz).
June 10: AWS released its 192-core Graviton5 server processors. Compared with the previous generation, they promise up to 25% more compute performance and up to 35% faster AI model execution; Meta will deploy them at a scale of “tens of millions of cores” (AWS).
June 12: France’s Mistral AI is in talks over an investment of around 3 billion euros at a valuation of about 20 billion euros, according to Bloomberg. In September, the company was valued at 11.7 billion euros (The Decoder).

Regulation and Law

The US government made Anthropic shut down Fable 5 and Mythos 5

On Friday, June 12, at 5:21 p.m. Eastern time, Anthropic received an order from the US government under export control rules. It barred all foreign nationals from accessing the Fable 5 and Mythos 5 models, including those living in the US and the company’s own employees (Anthropic).

Because the company cannot verify users’ citizenship in real time, it shut down both models for all customers worldwide. Other models, including Claude Opus 4.8, kept working (Fortune).

Reason: According to Axios, as cited by Nextgov, the Commerce Department stepped in after another company claimed it had managed to “jailbreak” the model, that is, bypass its safeguards (Nextgov).
Anthropic’s position: The company complied with the order but disagrees with it. In its view, this is a narrow, non-universal weakness, and similar results can be obtained from other public models, such as GPT-5.5.
If the same standard were applied across the industry, it “would essentially halt the deployment of all new models,” Anthropic said. In its view, the government should be able to block dangerous deployments, but through a legal process that is transparent and grounded in technical facts.
That same day, Anthropic also asked AWS to shut down both models in Amazon Bedrock (Info-Tech).

The case shows that the US government is starting to treat the most powerful models much like strategic technologies. For companies that build their services on new models, this means a new risk: a model they have paid for can disappear overnight.

Security and Incidents

June 12: The dispute over Fable 5 raised the question of how serious a discovered jailbreak must be to justify pulling a model. Anthropic said that in thousands of hours of pre-release testing, no one found a universal way to bypass the safeguards (Anthropic).
June 10: The Gemini outage once again showed how dependent businesses are on the availability of cloud AI services; business Google Workspace accounts were affected as well (9to5Google).