June

9 events

Editors' summary

Government halts a model for 19 days; both labs move toward IPO

On 9 June, Anthropic released Claude Fable 5 and Mythos 5 — one base model split in two by whether its safeguards were in place. On the 12th, the US government applied export controls to both. With no way to verify nationality in real time, Anthropic cut off every user worldwide. It was the first time the government had stopped the distribution of a specific model outright. The controls lifted on the 30th and the models returned after 19 days, with a new classifier added.

On the 1st, Anthropic filed a draft S-1 confidentially with the SEC; on the 8th, OpenAI disclosed the same step. Both revealed their IPO preparations in the same month.

On the 2nd, Microsoft released MAI-Thinking-1, its own reasoning model, at Build. Anthropic shipped Claude Sonnet 5 on the 30th.

This block is written by the editors. It is kept separate from the sourced record below.

Record9 events

1
BusinessAnthropic
Anthropic confidentially files a draft S-1 with the SEC
Anthropic disclosed that it had confidentially submitted a draft registration statement on Form S-1 to the US Securities and Exchange Commission, saying the number of shares and the offering price were not yet set and that timing would depend on market conditions. A confidential filing lets a company work through SEC feedback before the prospectus becomes public, and marks an early stage of listing preparation. It came less than a week after closing the $65 billion Series H at a $965 billion valuation, and put Anthropic ahead of OpenAI, which was readying a confidential filing of its own. Revenue that had run at about $10 billion a year in 2025 was at a $47 billion run rate as of May. Little more than five years after founding, a technology listing approaching $1 trillion was in view.
2
ProductAnthropic
Glasswing widens to 150 organisations, with 10,000 flaws found
Anthropic widened Project Glasswing to around 150 more organisations across more than fifteen countries, bringing in sectors that had been thinly represented — power, water, healthcare, communications, hardware. Many maintain codebases in wide use; a successful attack on one of them, the company estimated, could affect more than 100 million people. The fifty or so partners who had been in since April had by then run Claude Mythos Preview against their own code and found more than 10,000 high- or critical-severity flaws. It put a number on how far April's arrangement — withholding the model from general release and handing it only to defenders — had actually reached.
2
ModelMicrosoft / Mustafa Suleyman
Microsoft's first in-house reasoning model debuts at Build 2026
At Build 2026, Microsoft released MAI-Thinking-1, its first in-house reasoning model: 35 billion active parameters, trained from scratch on enterprise-grade commercially licensed data with no distillation from any third-party model. That last claim — not using another company's model outputs as a teacher — was itself a design decision shaped by the wave of litigation over where training data comes from. Alongside it came MAI-Code-1-Flash, a 5-billion-parameter coding model rolling out across all GitHub Copilot plans. Counting updated image, transcription, and voice models, Microsoft announced seven MAI models at Build 2026. Ten months after its first in-house models in August 2025, Microsoft had assembled its own reasoning, coding, voice, and image capability — turning the strategy of reducing dependence on OpenAI into an actual product lineup.
8
BusinessOpenAI / Sam Altman
OpenAI discloses confidential S-1 submission to SEC for potential IPO
OpenAI disclosed that it had confidentially submitted a draft S-1 registration statement to the SEC for a potential initial public offering. The company said it was announcing the submission itself because it expected the news to leak, and that timing was undecided and "may be a while." OpenAI was last valued at $852 billion post-money in its most recent funding round. Coming about a week after Anthropic was reported to have confidentially filed for its own IPO, the move was seen as the starting gun in a race of AI companies toward the public markets.
9
ModelAnthropic
Claude Fable 5 and Mythos 5: the same model, split by safeguards
Anthropic released two Mythos-class models, Claude Fable 5 and Claude Mythos 5, saying Fable 5's capabilities exceeded those of any model it had ever made generally available. The two shared the same underlying model and differed only in safeguards: the general-use Fable 5 automatically routed to Claude Opus 4.8 whenever classifiers for cybersecurity, biology and chemistry, or distillation fired, while Mythos 5 ran with those safeguards lifted and was restricted to authorized cybersecurity partners and, soon, selected biomedical researchers through Project Glasswing and a trusted access program. Stripe reported that Fable 5 "compressed months of engineering into days" on a 50-million-line codebase. Pricing was $10 and $50 per million input and output tokens. Shipping one model as two products divided by who is allowed to use it drew attention as a new template for handling frontier capability.
12
PolicyAnthropic
US export controls force Anthropic to pull Fable 5 worldwide
The US government applied export controls to Claude Fable 5 and Claude Mythos 5, and Anthropic suspended access to both for every user worldwide three days after launch. The directive barred access by foreign nationals, but with no way to verify nationality in real time the company shut the models off entirely; it said the directive arrived at 5:21pm ET that day. The government's basis was a method it believed it had found for bypassing Fable 5's safeguards. After reviewing a demonstration, Anthropic assessed it as a "narrow, non-universal jailbreak" involving asking the model to read code and identify flaws — capabilities it said were widely available from other models. The company objected that "if this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers." It was the first time the government directly halted distribution of a specific model.
24
ProductOpenAI / Broadcom
OpenAI and Broadcom unveil Jalapeño, OpenAI's first custom chip
OpenAI announced Jalapeño, an inference-only chip developed with Broadcom and the first custom silicon OpenAI has designed itself. The design targets inference rather than training, and OpenAI said its own models were used in the design work, taking the chip from conception to tape-out in about nine months. Initial deployment was planned for later in 2026, with volume production in 2027. With inference demand becoming the dominant cost, the chip marked a move to serve part of a compute base long dependent on NVIDIA GPUs with silicon of OpenAI's own.
30
ModelAnthropic
Claude Sonnet 5 closes in on Opus 4.8
Anthropic released Claude Sonnet 5, calling it its most agentic Sonnet yet, capable of planning, tool use, and operating autonomously, and saying that at higher effort levels it approached Opus 4.8 at lower cost. It improved substantially over Sonnet 4.6 in reasoning, tool use, coding, and knowledge work, with gains on agentic search (BrowseComp) and computer use (OSWorld-Verified). Introductory pricing of $2 and $10 per million input and output tokens was made permanent, and the model was available on every plan from Free to Enterprise. On safety, Anthropic reported a lower overall rate of undesirable behaviors than Sonnet 4.6, cyber safeguards on by default, and substantially less capability for developing software exploits than the Opus models. An updated tokenizer meant token counts could rise by 1.0 to 1.35x for the same text. The release coincided with the announcement lifting export controls on Fable 5.
30
PolicyAnthropic
Export controls lifted; Fable 5 returns after 19 days
Anthropic announced that the export controls on Claude Fable 5 and Mythos 5 had been lifted, with Fable 5 returning the next day, July 1, to users worldwide on the Claude Platform, claude.ai, Claude Code, and Claude Cowork — about 19 days after being pulled. The redeployment added a new safety classifier that blocked the jailbreak at issue in over 99% of cases and rerouted those requests to Claude Opus 4.8. Usage was metered at first, capped at up to 50% of weekly limits through July 7 and then handled through usage credits. Distribution resumed only after working with the government and layering on further technical mitigations — a demonstration that deploying a frontier model had become inseparable from government judgment.

↑ 2026