AI Bulletin · week of June 8

This week’s bulletin gathers six news blocks on AI applied to work, from 8 to 12 June 2026, plus a closing framework on the relationship between companies and the frontier labs. It covers model and agent launches, two regulatory decisions and an intense week of capital and infrastructure deals.

1. Anthropic launches Claude Fable 5 and Mythos 5, and reverses course on a set of invisible limits

Anthropic has announced Claude Fable 5 for general use and Claude Mythos 5 for a selected group of cyberdefenders and infrastructure providers. The company described both models as capable across software engineering, research, vision and cybersecurity, with more conservative safeguards applied to Fable 5. According to TLDR’s coverage, Fable 5 offers performance comparable to Mythos 5 with stricter safeguards and a one-million-token context window.

Alongside the launch, Anthropic introduced a set of safeguards that limit the model’s effectiveness in certain situations, such as when competing labs use Claude to develop models, through prompt modification, steering factors and parameter-efficient fine-tuning. These interventions were not visible to the user, and Fable 5 did not fall back to an alternative model. The company stated they would affect 0.03% of developers.

After criticism from researchers, who documented that Fable 5 was refusing or degrading responses on tasks such as training competing models, debugging AI code or optimising neural architectures, Anthropic decided to make those safeguards visible. Some reports noted that tokens and money had been spent on a model that did not behave as expected.

Sources: Anthropic — “Claude Fable 5 and Mythos 5”; Engadget, 12 June — “Anthropic walks back policy sabotaging research”; reports via TLDR AI, 10 and 12 June — “If Claude Fable stops helping you, you’ll never know” and “Anthropic backtracks on policy that ‘sabotaged’ researchers’ work”.

The bulletin mascot with a magnifying glass reviewing a freshly printed press sheet with several lines blacked out, on an oak newsroom desk — Anthropic launches Fable 5 and Mythos 5 and, after criticism, makes visible a set of safeguards that previously operated without telling the user.

2. Coding agents: OpenAI buys Ona, while Xiaomi and Cohere release open alternatives

OpenAI has announced the acquisition of Ona to bring secure cloud execution and orchestration capabilities to the Codex platform. According to the company, the technology is meant to support persistent, customer-controlled environments where agents can keep working across extended periods and sessions.

Xiaomi has released MiMo Code V0.1.0, an open-source, terminal-native coding assistant that, according to VentureBeat, outperforms Claude Code on several agentic coding benchmarks, particularly on long tasks of more than 200 steps. MiMo Code includes a cross-session memory system that uses an independent subagent to note decisions, issues and project scope as it progresses. It is available on GitHub under an MIT licence.

Cohere has launched North Mini Code, a mixture-of-experts coding model with 30 billion parameters and 3 billion active, released under an Apache 2.0 licence and aimed at sovereign AI environments. In parallel, Cursor updated its Bugbot reviewer to run more than three times faster, cost 22% less and find 10% more bugs per review, with most runs finishing in under three minutes.

Sources: report via TLDR AI, 12 June — “OpenAI acquired Ona for long-running agents”; VentureBeat — “Xiaomi’s new open source, agentic AI coding harness MiMo Code beats Claude Code at ultra-long, 200+ step tasks”; Cohere — “North Mini Code”; Cursor — “Bugbot updates June 2026”.

3. Inference speed: Xiaomi hits 1,000 tokens per second and Google speeds up Gemma with diffusion

Xiaomi, together with its inference partner TileRT, has introduced MiMo-V2.5-Pro-UltraSpeed, a one-trillion-parameter model with an inference speed of 1,000 tokens per second on a standard 8-GPU commodity node. The speed was achieved with FP4 quantisation on the expert layers and DFlash speculative decoding, which proposes a full block of tokens in a single pass instead of one at a time. The model is available through a limited API trial from 9 to 23 June and costs three times the standard MiMo-V2.5-Pro rate for roughly ten times the output. Decrypt described it as 15 times faster than ChatGPT and Claude.

Google has published DiffusionGemma, a mixture-of-experts model with 26 billion parameters that uses text diffusion to generate blocks of text simultaneously, with up to a fourfold speed increase on GPU. It is aimed at latency-sensitive applications, offers bidirectional attention and fits on high-end consumer GPUs when quantised, at the cost of some quality compared with standard models.

Sources: Decrypt — “China’s Xiaomi MiMo is now 15x faster than ChatGPT and Claude”; Google — “DiffusionGemma: 4x faster text generation”; reports via TLDR AI, 9 and 11 June.

The bulletin mascot next to a teletype machine spitting out a long blank paper ribbon at high speed, with motion lines and a stopwatch on the desk — The week brought several inference-speed announcements: 1,000 tokens per second at Xiaomi and up to fourfold faster generation in DiffusionGemma.

4. Consumer assistants: Apple unveils “Siri AI” and Google launches live voice translation

Apple has announced the Apple Intelligence update for Siri, now branded “Siri AI”, with a more conversational assistant planned for this fall’s operating-system releases. According to Ars Technica, the update includes Google-powered changes to the Foundation Models that Apple runs on the device and deeper AI integration across its platforms.

Google has introduced Gemini 3.5 Live Translate, an audio model for real-time speech-to-speech translation across more than 70 languages that removes pauses and keeps intonation. It is rolling out through Google products, including Meet in private preview and Google Translate on Android and iOS.

Sources: Ars Technica — “Say hi to Siri AI: Apple announces new, more conversational voice assistant”; Google — “Fluid, natural voice translation with Gemini 3.5 Live Translate”; reports via TLDR AI, 9 and 10 June.

5. Regulation: the EU orders Meta to open WhatsApp to rival chatbots, and Amodei proposes a framework for AI

The European Union has ordered Meta to open WhatsApp to chatbots from rival companies for free. Meta had banned third-party AI chatbots from the WhatsApp for Business API in October last year. The EU holds that Meta was abusing its dominant position in messaging in Europe by preventing competing assistants from using the WhatsApp API. Meta has announced it will appeal the decision, describing it as regulatory overreach that would grant some of the world’s largest companies access to the API without paying.

In parallel, Dario Amodei, Anthropic’s CEO, published the essay “Policy on the AI Exponential”, in which he argues for a regulatory approach to AI comparable to the FAA, with mandatory testing and stronger security standards. The text also addresses adapting macroeconomics and tax systems to AI-driven growth, regulatory reform to strengthen AI’s biomedical impact and aligning AI with democratic values on a global scale.

Sources: Engadget — “EU orders Meta to stop blocking rival AI chatbots on WhatsApp”; Dario Amodei — “Policy on the AI Exponential”; reports via TLDR AI, 11 June.

6. Money and infrastructure: OpenAI files its S-1, Google backstops Anthropic’s chips and Oracle drops

OpenAI has stated that it confidentially filed a draft S-1 registration statement with the SEC, with no timing decided for a possible IPO; according to the company, the filing preserves the option to list sooner while it continues weighing whether to remain private. Perplexity, for its part, has placed its IPO in 2028 regardless of what Anthropic or OpenAI do.

Google is backstopping Anthropic’s $35 billion chip lease by guaranteeing payments at five data centres, an Anthropic role in that financing that had not been disclosed before. OpenAI, in addition, is weighing an Nvidia-backed lease for a 10 GW data-centre campus in Ohio, under a 20-year contract and expected to begin operating in 2028.

Oracle fell 11% on the stock market after telling investors to expect an additional $20 billion capital raise and reporting negative free cash flow for the year. The company grew revenue in its fiscal fourth quarter, but its AI buildout pushed capital expenditures up 162%, to $55.7 billion.

Sources: report via TLDR AI, 9 June — “OpenAI filed a confidential S-1”; CNBC — “Oracle shares tumble 11% on increased capital raise, cash concerns”; reports via TLDR AI, 10 and 11 June — “Google’s backstops underpin $35 billion chip deal for Anthropic” and “OpenAI weighs Nvidia-backed lease for 10 GW Ohio data center campus”.

The bulletin mascot at a newsroom desk surrounded by stacked newspapers and a blank ticker tape flowing from a machine, with a coffee cup and a monstera plant in the background — A busy week of capital and infrastructure: OpenAI's confidential S-1, Google's backstop of Anthropic's chip lease and Oracle's drop on the stock market.

7. Framework of the week: the relationship between companies and the frontier labs

Several pieces that circulated this week point to how the relationship between the companies that buy AI and the labs that provide it is being strained.

A statement from Palantir’s CEO, Alex Karp, carried by CNBC, holds that its enterprise customers are unhappy with how the frontier labs operate and that the labs only care about burning tokens to signal productivity.
The Fable 5 safeguards episode led several analysts to describe a supply-chain risk: a company cannot know whether it is hitting those limits, which undermines trust in the tools.
An analysis titled “Moats need models” argues that defensibility comes from owning the full loop of model, harness, workflow and evaluation, rather than renting frontier capability that the provider can restrict, reprice or reclaim.
A thread on labs’ business model notes that the gross margins on subscriptions are worse than on the API, so labs are likely to start reserving new features or models for usage-based plans.
AI Gateway data cited this week puts the jump in DeepSeek’s token share from under 1% to 17% in a month, while its share of spend stayed near 1%.

Sources: CNBC — “Palantir’s Karp says businesses are ‘unhappy’ with the frontier AI labs”; reports via TLDR AI, 10, 11 and 12 June — “Claude Fable 5 and new AI safety fables”, “Moats need models”, “What’s the better business model for an AI lab, subscription or API?” and “DeepSeek enters the fight for token volume”.

Closing

The six blocks above cover the launch of Claude Fable 5 and Mythos 5 alongside Anthropic’s reversal on a set of invisible safeguards, a batch of coding agents and models from OpenAI, Xiaomi and Cohere, the advances in inference speed from Xiaomi and Google, the consumer-assistant updates from Apple and Google, two regulatory moves in the EU and in Anthropic’s policy discourse, and a round of capital and infrastructure deals involving OpenAI, Google, Anthropic and Oracle. The framework of the week gathers five readings on how the relationship between companies and the frontier labs is being strained. The items reflect what the companies and the cited reports have communicated; they do not include productivity data verified in specific organisations.