AI Product Updates Daily — June 13, 2026

The US government shut down Anthropic's most advanced models three days after launch. OpenAI got subpoenaed by a coalition of state attorneys general. SpaceX's SPCX triggered mandatory MSCI index buying on day two. And Anthropic quietly overtook OpenAI in paid US business adoption for the first time. That's your Friday — June 13, 2026.

Anthropic's Fable 5 and Mythos 5 pulled by US export control order

Three days after launching Claude Fable 5 and Mythos 5, Anthropic disabled both models for every customer worldwide. The order arrived at 5:21 PM ET on June 12 from Commerce Secretary Howard Lutnick, citing national security authorities. It banned access by any foreign national — including Anthropic's own foreign-national employees — and Anthropic said it had no way to filter users in real time, so it shut down access for everyone.1

The government's stated concern: another company had reportedly jailbroken Mythos 5, alarming officials about potential security risks. Anthropic's own review found the technique narrow and non-universal — essentially asking the model to read a codebase and identify vulnerabilities — and said the same capability was available from other public models including OpenAI's GPT-5.5.2 Anthropic published its rebuttal openly:

"We believe the government should have the ability to block unsafe deployments, as part of a statutory process that is transparent, fair, clear, and grounded in technical facts. This action does not adhere to those principles."

It called the situation a likely misunderstanding and said it was working to restore access. All other Claude models — Opus 4.8, Sonnet, Haiku — remain fully available.

Model	Status	Pricing	Notes
Claude Fable 5	Disabled for all users	$10/M in, $50/M out	Mythos-class, was public via `claude-fable-5`
Claude Mythos 5	Disabled for all users	$10/M in, $50/M out	Project Glasswing partners only, cyber safeguards lifted
Claude Opus 4.8	Fully available	Standard rates	Recommended fallback for builders
Claude Sonnet / Haiku	Fully available	Standard rates	Unaffected

This appears to be the first government-forced takedown of a publicly deployed frontier model.3 Developers who built on the claude-fable-5 API string will see request errors until access is restored. Anthropic's earlier clash with the DOD — which declared the company a supply chain risk — is still in litigation.

OpenAI subpoenaed by coalition of state attorneys general

A coalition of state attorneys general opened an investigation into OpenAI on Friday, serving the company with a subpoena seeking documents on advertising practices, user engagement and retention, handling of consumer and health data, activities related to minors and seniors, model sycophancy, and company policies.4 The subpoena was sent by New York's attorney general.

The scope spans OpenAI's commercial practices broadly, not a single incident. The WSJ report noted that the move is "the latest in a series of legal actions by states directed at AI companies." It follows a separate lawsuit claiming OpenAI and Sam Altman knowingly released an unsafe product.

SpaceX SPCX: MSCI index buying begins on day two

SpaceX's Nasdaq debut closed June 12 up roughly 19–25 percent from its $135 IPO price (reporting varied between $161.11 and $168.70 on different tracking sources), giving it a market cap above $2 trillion and confirming it as the largest IPO in history.5 On June 13 — today — MSCI began applying early inclusion rules for large IPOs, meaning passive index funds tracking MSCI benchmarks started mandatory SPCX purchases.6 Nasdaq also adjusted its rules to allow fast-track inclusion into the Nasdaq-100 after 15 trading days, placing that structural buying event around early July.7

正在加载图表…

The xAI segment, which includes Grok and the Colossus data centers (acquired in February 2026), posted a $6.36 billion operating loss in 2025. The S-1 disclosed a cumulative loss of $41.3 billion since SpaceX's founding in 2002, though Starlink's 2025 EBITDA margin of roughly 63% on $11.4 billion in revenue is the profitability story underwriting the IPO valuation.

For AI developers: xAI, now a public company, has its financial disclosures visible in quarterly filings.

Anthropic overtakes OpenAI in paid US business adoption

The May 2026 Ramp AI Index — tracking spending by more than 50,000 US businesses — showed Anthropic's adoption reaching 34.4% in April, up 3.8 percentage points, while OpenAI slipped to 32.3%, down 2.9 points.8 That's the first time Anthropic has led on this metric since tracking began.

Ramp AI Index — April 2026

US business AI adoption — corporate card spending across 50,000+ businesses

Anthropic adoption

34.4%

OpenAI adoption

32.3%

Overall AI adoption

50.6%

GitHub commits by Claude Code (est.)

正在加载统计卡片…

The engine: Claude Code. One analysis estimated that 4% of all public GitHub commits worldwide are now authored by Claude Code — double the share from a month earlier.

A separate IDC survey of 1,000+ organizations complicates the picture. It found only 19% reporting extensive use of Claude, compared to higher figures for OpenAI and Google. The gap between the two datasets is real: Ramp measures which vendor a business is paying for (a binary signal), while IDC measures depth of usage and can reflect incumbent advantages from OpenAI's longer market presence and Google's Workspace bundling.

Meanwhile, The Information reported that Anthropic has been surprising business partners with competitive product launches and unannounced pricing changes — including using Figma and Canva as "launch partners" for Claude Design while that tool competes directly with their products.9 For enterprise buyers weighing multi-year contracts ahead of an Anthropic IPO, partner trust is a revenue-durability question.

Claude Code session branching — the new /fork command allows developers to split a session into parallel variants without losing the original state. — Claude Code /fork command introduced June 13 — branches sessions the way git branches code. 3

Claude Code gets /fork command and new CLI

Separately from the Fable 5 shutdown, Anthropic shipped two Claude Code updates on June 13: a /fork command that branches an existing session into a parallel variant (useful for testing different approaches without losing the original session state), and a new command-line interface.9 This follows the earlier v2.1.173 update (covered June 12) that added nested sub-agents and model-name normalization. The /fork pattern maps directly onto how developers already use git branches, lowering the learning curve for teams already comfortable with version control workflows.

Z.ai ships GLM-5.2 on the GLM Coding Plan

Z.ai released GLM-5.2 today as its new flagship model, available immediately to every GLM Coding Plan tier — Lite, Pro, Max, and Team. The company claims 1M-token context (exposed via the glm-5.2[1m] model ID), two thinking-effort levels (High and Max), and continued strength on long-horizon agentic tasks.10

What's not available yet: standalone API access, the Z.ai chatbot, and open-source weights (all announced for next week under MIT License). No benchmarks were published at launch. The GLM-5 family has a credible track record — GLM-5 scored 77.8% on SWE-bench Verified, and GLM-5.1 reportedly reached roughly 94.6% of Claude Opus 4.6's coding score in internal benchmarks — but those numbers belong to earlier models, not GLM-5.2 specifically. Independent testing won't happen until the API and weights ship.

GLM Coding Plan tier	5-hour prompt cap	Weekly prompt cap
Lite	~80	~400
Pro	~400	~2,000
Max	~1,600	~8,000
Team	Seat-based	See Z.ai

For teams on the GLM Coding Plan, switching is an environment variable change: set ANTHROPIC_DEFAULT_SONNET_MODEL to glm-5.2[1m] in ~/.claude/settings.json.

Google Home adds four Gemini-powered features

Google pushed four updates to Google Home, available to early access members since June 8:11

Hourly weather forecasts with improved speech accuracy and temperature-unit consistency; hourly charts on smart displays
Gemini-driven video discovery — conversationally search and launch movies and TV shows from connected streaming services
Direct streaming from connected streaming services (Netflix, etc.) without using a phone or separate remote
Improved noise handling — background noise rejection, better natural-language understanding for volume and reminder management

No specific timeline was given for the wider rollout beyond early access. This follows the broader Gemini integration into Google Home hardware that Google has been building since late 2025.

OpenAI: GPT-5.6 leak and multi-state AG investigation

Two separate OpenAI threads on Friday:

An unreleased checkpoint identified as GPT-5.6 "kindle-alpha" surfaced in developer discussions, apparently through Codex testing paths, with reports of stronger reasoning, coding, and vision performance than earlier GPT-5.x models. Leaker Jimmy Apples noted on X that GPT-5.5 was already improving rapidly day-over-day, and the new checkpoint appeared to extend that trajectory.12 Prior codenames "ember-alpha" and "beacon-alpha" also appeared in rollout logs, with some ChatGPT Pro users reporting context behavior consistent with a 1.5M token window — 43% above GPT-5.5's documented limit.9 OpenAI has made no official announcement. A leaked checkpoint name is not a model card; treat GPT-5.6 as a watchlist item, not a roadmap input.

On the legal front, a coalition of state AGs subpoenaed OpenAI (covered above), while Elon Musk's megatrial against OpenAI and Sam Altman continued in its second week, with scrutiny on an OpenAI executive's finances.4

Quick signals

Mammouth AI updated its supported model list, adding Claude Opus 4.8, Gemini 3.5 Flash, Mistral Medium 3.5, and Grok 4.3 on June 12.13
EngineAI, the Shenzhen humanoid robot maker, filed confidentially for a Hong Kong IPO. Its 12,000 sq meter Shenzhen factory opened June 1 with a line producing one T800 humanoid robot every 15 minutes, targeting 10,000 units.9
Microsoft Build 2026 coverage continued, with MAI Thinking One (reasoning, coding, image generation) and Alibaba's Qwen 3.7 Plus (multimodal, vision + language + coding) both drawing comparisons as the June model-release compression continues.
Anthropic's Amazon compute deal expanded to up to 5 gigawatts of training and inference capacity, per TechWire Asia's summary of Anthropic's first five months of 2026. Run-rate revenue above $30B, with 1,000+ customers spending more than $1M annually.